This is page iii
Printer: Opaque this
Jorge Nocedal
Stephen J. Wright
Numerical Optimization
Second Edition
This is pag
Printer: O
Jorge Nocedal
EECS Department
Northwestern University
Evanston, IL 60208-3118
USA
nocedal@eecs.northwestern.edu
Stephen J. Wright
Computer Sciences Department
University of Wisconsin
1210 West Dayton Street
Madison, WI 53706–1613
USA
swright@cs.wisc.edu
Series Editors:
Thomas V. Mikosch
University of Copenhagen
Laboratory of Actuarial Mathematics
DK-1017 Copenhagen
Denmark
mikosch@act.ku.dk
Stephen M. Robinson
Department of Industrial and Systems
Engineering
University of Wisconsin
1513 University Avenue
Madison, WI 53706–1539
USA
smrobins@facstaff.wise.edu
Sidney I. Resnick
Cornell University
School of Operations Research and
Industrial Engineering
Ithaca, NY 14853
USA
sirl@cornell.edu
Mathematics Subject Classification (2000): 90B30, 90C11, 90-01, 90-02
Library of Congress Control Number: 2006923897
ISBN-10: 0-387-30303-0
ISBN-13: 978-0387-30303-1
Printed on acid-free paper.
C 2006 Springer Science+Business Media, LLC.
All rights reserved. This work may not be translated or copied in whole or in part without the written permission
of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for
brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not
identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary
rights.
Printed in the United States of America.
9 8 7 6 5 4 3 2 1
springer.com
(TB/HAM)
This is page v
Printer: Opaque this
To Sue, Isabel and Martin
and
To Mum and Dad
This is page vii
Printer: Opaque this
Contents
Preface
xvii
Preface to the Second Edition
1
2
Introduction
Mathematical Formulation . . . . . . . . . .
Example: A Transportation Problem . . . . .
Continuous versus Discrete Optimization . . .
Constrained and Unconstrained Optimization
Global and Local Optimization . . . . . . . .
Stochastic and Deterministic Optimization . .
Convexity . . . . . . . . . . . . . . . . . . .
Optimization Algorithms . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . .
xxi
.
.
.
.
.
.
.
.
.
1
2
4
5
6
6
7
7
8
9
Fundamentals of Unconstrained Optimization
2.1
What Is a Solution? . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
12
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
viii
CONTENTS
Recognizing a Local Minimum . . . . . . .
Nonsmooth Problems . . . . . . . . . . . .
2.2
Overview of Algorithms . . . . . . . . . . .
Two Strategies: Line Search and Trust Region
Search Directions for Line Search Methods .
Models for Trust-Region Methods . . . . . .
Scaling . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . .
3
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
14
17
18
19
20
25
26
27
Line Search Methods
3.1
Step Length . . . . . . . . . . . . . . . . . . . .
The Wolfe Conditions . . . . . . . . . . . . . . .
The Goldstein Conditions . . . . . . . . . . . . .
Sufficient Decrease and Backtracking . . . . . . .
3.2
Convergence of Line Search Methods . . . . . . .
3.3
Rate of Convergence . . . . . . . . . . . . . . . .
Convergence Rate of Steepest Descent . . . . . . .
Newton’s Method . . . . . . . . . . . . . . . . .
Quasi-Newton Methods . . . . . . . . . . . . . .
3.4
Newton’s Method with Hessian Modification . . .
Eigenvalue Modification . . . . . . . . . . . . . .
Adding a Multiple of the Identity . . . . . . . . .
Modified Cholesky Factorization . . . . . . . . .
Modified Symmetric Indefinite Factorization . . .
3.5
Step-Length Selection Algorithms . . . . . . . . .
Interpolation . . . . . . . . . . . . . . . . . . . .
Initial Step Length . . . . . . . . . . . . . . . . .
A Line Search Algorithm for the Wolfe Conditions
Notes and References . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
30
31
33
36
37
37
41
42
44
46
48
49
51
52
54
56
57
59
60
62
63
Trust-Region Methods
Outline of the Trust-Region Approach . .
4.1
Algorithms Based on the Cauchy Point . .
The Cauchy Point . . . . . . . . . . . . .
Improving on the Cauchy Point . . . . . .
The Dogleg Method . . . . . . . . . . . .
Two-Dimensional Subspace Minimization
4.2
Global Convergence . . . . . . . . . . . .
Reduction Obtained by the Cauchy Point .
Convergence to Stationary Points . . . . .
4.3
Iterative Solution of the Subproblem . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
66
68
71
71
73
73
76
77
77
79
83
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
The Hard Case . . . . . . . . . . . . . . . . . . . . . . . .
Proof of Theorem 4.1 . . . . . . . . . . . . . . . . . . . .
Convergence of Algorithms Based on Nearly Exact Solutions
4.4
Local Convergence of Trust-Region Newton Methods . . .
4.5
Other Enhancements . . . . . . . . . . . . . . . . . . . .
Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Trust Regions in Other Norms . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
6
Conjugate Gradient Methods
5.1
The Linear Conjugate Gradient Method . . . . . . .
Conjugate Direction Methods . . . . . . . . . . . .
Basic Properties of the Conjugate Gradient Method
A Practical Form of the Conjugate Gradient Method
Rate of Convergence . . . . . . . . . . . . . . . . .
Preconditioning . . . . . . . . . . . . . . . . . . .
Practical Preconditioners . . . . . . . . . . . . . .
5.2
Nonlinear Conjugate Gradient Methods . . . . . .
The Fletcher–Reeves Method . . . . . . . . . . . .
The Polak–Ribière Method and Variants . . . . . .
Quadratic Termination and Restarts . . . . . . . . .
Behavior of the Fletcher–Reeves Method . . . . . .
Global Convergence . . . . . . . . . . . . . . . . .
Numerical Performance . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quasi-Newton Methods
6.1
The BFGS Method . . . . . . . . . . . . . . .
Properties of the BFGS Method . . . . . . . .
Implementation . . . . . . . . . . . . . . . .
6.2
The SR1 Method . . . . . . . . . . . . . . . .
Properties of SR1 Updating . . . . . . . . . .
6.3
The Broyden Class . . . . . . . . . . . . . . .
6.4
Convergence Analysis . . . . . . . . . . . . .
Global Convergence of the BFGS Method . . .
Superlinear Convergence of the BFGS Method
Convergence Analysis of the SR1 Method . . .
Notes and References . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
87
89
91
92
95
95
97
98
98
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
101
102
102
107
111
112
118
120
121
121
122
124
125
127
131
132
133
.
.
.
.
.
.
.
.
.
.
.
.
135
136
141
142
144
147
149
153
153
156
160
161
162
ix
x
CONTENTS
7
8
9
Large-Scale Unconstrained Optimization
7.1
Inexact Newton Methods . . . . . . . . . . . . . . . .
Local Convergence of Inexact Newton Methods . . . . .
Line Search Newton–CG Method . . . . . . . . . . . .
Trust-Region Newton–CG Method . . . . . . . . . . .
Preconditioning the Trust-Region Newton–CG Method
Trust-Region Newton–Lanczos Method . . . . . . . . .
7.2
Limited-Memory Quasi-Newton Methods . . . . . . .
Limited-Memory BFGS . . . . . . . . . . . . . . . . .
Relationship with Conjugate Gradient Methods . . . .
General Limited-Memory Updating . . . . . . . . . . .
Compact Representation of BFGS Updating . . . . . .
Unrolling the Update . . . . . . . . . . . . . . . . . .
7.3
Sparse Quasi-Newton Updates . . . . . . . . . . . . .
7.4
Algorithms for Partially Separable Functions . . . . . .
7.5
Perspectives and Software . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
164
165
166
168
170
174
175
176
177
180
181
181
184
185
186
189
190
191
Calculating Derivatives
8.1
Finite-Difference Derivative Approximations
Approximating the Gradient . . . . . . . . .
Approximating a Sparse Jacobian . . . . . .
Approximating the Hessian . . . . . . . . .
Approximating a Sparse Hessian . . . . . . .
8.2
Automatic Differentiation . . . . . . . . . .
An Example . . . . . . . . . . . . . . . . .
The Forward Mode . . . . . . . . . . . . .
The Reverse Mode . . . . . . . . . . . . . .
Vector Functions and Partial Separability . .
Calculating Jacobians of Vector Functions . .
Calculating Hessians: Forward Mode . . . .
Calculating Hessians: Reverse Mode . . . . .
Current Limitations . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
193
194
195
197
201
202
204
205
206
207
210
212
213
215
216
217
217
Derivative-Free Optimization
9.1
Finite Differences and Noise . . . .
9.2
Model-Based Methods . . . . . . .
Interpolation and Polynomial Bases
Updating the Interpolation Set . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
220
221
223
226
227
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
A Method Based on Minimum-Change Updating .
Coordinate and Pattern-Search Methods . . . . .
Coordinate Search Method . . . . . . . . . . . .
Pattern-Search Methods . . . . . . . . . . . . . .
9.4
A Conjugate-Direction Method . . . . . . . . . .
9.5
Nelder–Mead Method . . . . . . . . . . . . . . .
9.6
Implicit Filtering . . . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
9.3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
228
229
230
231
234
238
240
242
242
10 Least-Squares Problems
10.1 Background . . . . . . . . . . . . . . . . . . . . . .
10.2 Linear Least-Squares Problems . . . . . . . . . . . .
10.3 Algorithms for Nonlinear Least-Squares Problems . .
The Gauss–Newton Method . . . . . . . . . . . . . .
Convergence of the Gauss–Newton Method . . . . . .
The Levenberg–Marquardt Method . . . . . . . . . .
Implementation of the Levenberg–Marquardt Method
Convergence of the Levenberg–Marquardt Method . .
Methods for Large-Residual Problems . . . . . . . . .
10.4 Orthogonal Distance Regression . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
245
247
250
254
254
255
258
259
261
262
265
267
269
11 Nonlinear Equations
11.1 Local Algorithms . . . . . . . . . . . . . .
Newton’s Method for Nonlinear Equations
Inexact Newton Methods . . . . . . . . .
Broyden’s Method . . . . . . . . . . . . .
Tensor Methods . . . . . . . . . . . . . .
11.2 Practical Methods . . . . . . . . . . . . .
Merit Functions . . . . . . . . . . . . . .
Line Search Methods . . . . . . . . . . . .
Trust-Region Methods . . . . . . . . . . .
11.3 Continuation/Homotopy Methods . . . .
Motivation . . . . . . . . . . . . . . . . .
Practical Continuation Methods . . . . . .
Notes and References . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
270
274
274
277
279
283
285
285
287
290
296
296
297
302
302
12 Theory of Constrained Optimization
Local and Global Solutions . . . . . . . . . . . . . . . . . . . . . . . .
304
305
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xi
xii
CONTENTS
Smoothness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A Single Equality Constraint . . . . . . . . . . . . . . . . . . . . . .
A Single Inequality Constraint . . . . . . . . . . . . . . . . . . . . .
Two Inequality Constraints . . . . . . . . . . . . . . . . . . . . . .
12.2 Tangent Cone and Constraint Qualifications . . . . . . . . . . . . .
12.3 First-Order Optimality Conditions . . . . . . . . . . . . . . . . . .
12.4 First-Order Optimality Conditions: Proof . . . . . . . . . . . . . . .
Relating the Tangent Cone and the First-Order Feasible Direction Set
A Fundamental Necessary Condition . . . . . . . . . . . . . . . . .
Farkas’ Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Proof of Theorem 12.1 . . . . . . . . . . . . . . . . . . . . . . . . .
12.5 Second-Order Conditions . . . . . . . . . . . . . . . . . . . . . . .
Second-Order Conditions and Projected Hessians . . . . . . . . . .
12.6 Other Constraint Qualifications . . . . . . . . . . . . . . . . . . . .
12.7 A Geometric Viewpoint . . . . . . . . . . . . . . . . . . . . . . . .
12.8 Lagrange Multipliers and Sensitivity . . . . . . . . . . . . . . . . . .
12.9 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.1
13 Linear Programming: The Simplex Method
Linear Programming . . . . . . . . . . . .
13.1 Optimality and Duality . . . . . . . . . .
Optimality Conditions . . . . . . . . . . .
The Dual Problem . . . . . . . . . . . . .
13.2 Geometry of the Feasible Set . . . . . . . .
Bases and Basic Feasible Points . . . . . .
Vertices of the Feasible Polytope . . . . . .
13.3 The Simplex Method . . . . . . . . . . . .
Outline . . . . . . . . . . . . . . . . . . .
A Single Step of the Method . . . . . . . .
13.4 Linear Algebra in the Simplex Method . .
13.5 Other Important Details . . . . . . . . . .
Pricing and Selection of the Entering Index
Starting the Simplex Method . . . . . . .
Degenerate Steps and Cycling . . . . . . .
13.6 The Dual Simplex Method . . . . . . . . .
13.7 Presolving . . . . . . . . . . . . . . . . .
13.8 Where Does the Simplex Method Fit? . . .
Notes and References . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
306
307
308
310
313
315
320
323
323
325
326
329
330
337
338
340
341
343
349
351
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
355
356
358
358
359
362
362
365
366
366
370
372
375
375
378
381
382
385
388
389
389
CONTENTS
14 Linear Programming: Interior-Point Methods
14.1 Primal-Dual Methods . . . . . . . . . . . . . . . . . . .
Outline . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Central Path . . . . . . . . . . . . . . . . . . . . . .
Central Path Neighborhoods and Path-Following Methods
14.2 Practical Primal-Dual Algorithms . . . . . . . . . . . . .
Corrector and Centering Steps . . . . . . . . . . . . . . .
Step Lengths . . . . . . . . . . . . . . . . . . . . . . . .
Starting Point . . . . . . . . . . . . . . . . . . . . . . .
A Practical Algorithm . . . . . . . . . . . . . . . . . . .
Solving the Linear Systems . . . . . . . . . . . . . . . . .
14.3 Other Primal-Dual Algorithms and Extensions . . . . . .
Other Path-Following Methods . . . . . . . . . . . . . .
Potential-Reduction Methods . . . . . . . . . . . . . . .
Extensions . . . . . . . . . . . . . . . . . . . . . . . . .
14.4 Perspectives and Software . . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
392
393
393
397
399
407
407
409
410
411
411
413
413
414
415
416
417
418
15 Fundamentals of Algorithms for Nonlinear Constrained Optimization
15.1 Categorizing Optimization Algorithms . . . . . . . . . . . . . .
15.2 The Combinatorial Difficulty of Inequality-Constrained Problems
15.3 Elimination of Variables . . . . . . . . . . . . . . . . . . . . . .
Simple Elimination using Linear Constraints . . . . . . . . . . .
General Reduction Strategies for Linear Constraints . . . . . . .
Effect of Inequality Constraints . . . . . . . . . . . . . . . . . .
15.4 Merit Functions and Filters . . . . . . . . . . . . . . . . . . . .
Merit Functions . . . . . . . . . . . . . . . . . . . . . . . . . .
Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.5 The Maratos Effect . . . . . . . . . . . . . . . . . . . . . . . . .
15.6 Second-Order Correction and Nonmonotone Techniques . . . .
Nonmonotone (Watchdog) Strategy . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
421
422
424
426
428
431
434
435
435
437
440
443
444
446
446
16 Quadratic Programming
16.1 Equality-Constrained Quadratic Programs
Properties of Equality-Constrained QPs . .
16.2 Direct Solution of the KKT System . . . .
Factoring the Full KKT System . . . . . .
Schur-Complement Method . . . . . . . .
Null-Space Method . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
448
451
451
454
454
455
457
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xiii
xiv
CONTENTS
16.3
Iterative Solution of the KKT System . . . . . . . . . . . . . . . .
CG Applied to the Reduced System . . . . . . . . . . . . . . . . .
The Projected CG Method . . . . . . . . . . . . . . . . . . . . . .
16.4 Inequality-Constrained Problems . . . . . . . . . . . . . . . . . .
Optimality Conditions for Inequality-Constrained Problems . . . .
Degeneracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16.5 Active-Set Methods for Convex QPs . . . . . . . . . . . . . . . . .
Specification of the Active-Set Method for Convex QP . . . . . . .
Further Remarks on the Active-Set Method . . . . . . . . . . . . .
Finite Termination of Active-Set Algorithm on Strictly Convex QPs
Updating Factorizations . . . . . . . . . . . . . . . . . . . . . . .
16.6 Interior-Point Methods . . . . . . . . . . . . . . . . . . . . . . .
Solving the Primal-Dual System . . . . . . . . . . . . . . . . . . .
Step Length Selection . . . . . . . . . . . . . . . . . . . . . . . .
A Practical Primal-Dual Method . . . . . . . . . . . . . . . . . .
16.7 The Gradient Projection Method . . . . . . . . . . . . . . . . . .
Cauchy Point Computation . . . . . . . . . . . . . . . . . . . . .
Subspace Minimization . . . . . . . . . . . . . . . . . . . . . . .
16.8 Perspectives and Software . . . . . . . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17 Penalty and Augmented Lagrangian Methods
17.1 The Quadratic Penalty Method . . . . . . . . . . . .
Motivation . . . . . . . . . . . . . . . . . . . . . . .
Algorithmic Framework . . . . . . . . . . . . . . . .
Convergence of the Quadratic Penalty Method . . . .
Ill Conditioning and Reformulations . . . . . . . . .
17.2 Nonsmooth Penalty Functions . . . . . . . . . . . .
A Practical 1 Penalty Method . . . . . . . . . . . . .
A General Class of Nonsmooth Penalty Methods . . .
17.3 Augmented Lagrangian Method: Equality Constraints
Motivation and Algorithmic Framework . . . . . . .
Properties of the Augmented Lagrangian . . . . . . .
17.4 Practical Augmented Lagrangian Methods . . . . . .
Bound-Constrained Formulation . . . . . . . . . . .
Linearly Constrained Formulation . . . . . . . . . .
Unconstrained Formulation . . . . . . . . . . . . . .
17.5 Perspectives and Software . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
459
459
461
463
464
465
467
472
476
477
478
480
482
483
484
485
486
488
490
492
492
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
497
498
498
501
502
505
507
511
513
514
514
517
519
519
522
523
525
526
527
CONTENTS
18 Sequential Quadratic Programming
18.1 Local SQP Method . . . . . . . . . . . . . . . . . . . . . . .
SQP Framework . . . . . . . . . . . . . . . . . . . . . . . .
Inequality Constraints . . . . . . . . . . . . . . . . . . . . .
18.2 Preview of Practical SQP Methods . . . . . . . . . . . . . . .
IQP and EQP . . . . . . . . . . . . . . . . . . . . . . . . .
Enforcing Convergence . . . . . . . . . . . . . . . . . . . .
18.3 Algorithmic Development . . . . . . . . . . . . . . . . . . .
Handling Inconsistent Linearizations . . . . . . . . . . . . .
Full Quasi-Newton Approximations . . . . . . . . . . . . . .
Reduced-Hessian Quasi-Newton Approximations . . . . . .
Merit Functions . . . . . . . . . . . . . . . . . . . . . . . .
Second-Order Correction . . . . . . . . . . . . . . . . . . .
18.4 A Practical Line Search SQP Method . . . . . . . . . . . . .
18.5 Trust-Region SQP Methods . . . . . . . . . . . . . . . . . .
A Relaxation Method for Equality-Constrained Optimization
S1 QP (Sequential 1 Quadratic Programming) . . . . . . .
Sequential Linear-Quadratic Programming (SLQP) . . . . .
A Technique for Updating the Penalty Parameter . . . . . . .
18.6 Nonlinear Gradient Projection . . . . . . . . . . . . . . . .
18.7 Convergence Analysis . . . . . . . . . . . . . . . . . . . . .
Rate of Convergence . . . . . . . . . . . . . . . . . . . . . .
18.8 Perspectives and Software . . . . . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19 Interior-Point Methods for Nonlinear Programming
19.1 Two Interpretations . . . . . . . . . . . . . . . . .
19.2 A Basic Interior-Point Algorithm . . . . . . . . . .
19.3 Algorithmic Development . . . . . . . . . . . . . .
Primal vs. Primal-Dual System . . . . . . . . . . .
Solving the Primal-Dual System . . . . . . . . . . .
Updating the Barrier Parameter . . . . . . . . . . .
Handling Nonconvexity and Singularity . . . . . . .
Step Acceptance: Merit Functions and Filters . . . .
Quasi-Newton Approximations . . . . . . . . . . .
Feasible Interior-Point Methods . . . . . . . . . . .
19.4 A Line Search Interior-Point Method . . . . . . . .
19.5 A Trust-Region Interior-Point Method . . . . . . .
An Algorithm for Solving the Barrier Problem . . .
Step Computation . . . . . . . . . . . . . . . . . .
Lagrange Multipliers Estimates and Step Acceptance
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
529
530
531
532
533
533
534
535
535
536
538
540
543
545
546
547
549
551
553
554
556
557
560
561
561
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
563
564
566
569
570
570
572
573
575
575
576
577
578
578
580
581
xv
xvi
CONTENTS
Description of a Trust-Region Interior-Point Method .
The Primal Log-Barrier Method . . . . . . . . . . . .
Global Convergence Properties . . . . . . . . . . . .
Failure of the Line Search Approach . . . . . . . . . .
Modified Line Search Methods . . . . . . . . . . . .
Global Convergence of the Trust-Region Approach . .
19.8 Superlinear Convergence . . . . . . . . . . . . . . .
19.9 Perspectives and Software . . . . . . . . . . . . . . .
Notes and References . . . . . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19.6
19.7
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A Background Material
A.1 Elements of Linear Algebra . . . . . . . . . . . . . . . . . . . .
Vectors and Matrices . . . . . . . . . . . . . . . . . . . . . . . .
Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Eigenvalues, Eigenvectors, and the Singular-Value Decomposition
Determinant and Trace . . . . . . . . . . . . . . . . . . . . . .
Matrix Factorizations: Cholesky, LU, QR . . . . . . . . . . . . .
Symmetric Indefinite Factorization . . . . . . . . . . . . . . . .
Sherman–Morrison–Woodbury Formula . . . . . . . . . . . . .
Interlacing Eigenvalue Theorem . . . . . . . . . . . . . . . . . .
Error Analysis and Floating-Point Arithmetic . . . . . . . . . . .
Conditioning and Stability . . . . . . . . . . . . . . . . . . . . .
A.2 Elements of Analysis, Geometry, Topology . . . . . . . . . . . .
Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Rates of Convergence . . . . . . . . . . . . . . . . . . . . . . .
Topology of the Euclidean Space IRn . . . . . . . . . . . . . . . .
Convex Sets in IRn . . . . . . . . . . . . . . . . . . . . . . . . .
Continuity and Limits . . . . . . . . . . . . . . . . . . . . . . .
Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Directional Derivatives . . . . . . . . . . . . . . . . . . . . . .
Mean Value Theorem . . . . . . . . . . . . . . . . . . . . . . .
Implicit Function Theorem . . . . . . . . . . . . . . . . . . . .
Order Notation . . . . . . . . . . . . . . . . . . . . . . . . . .
Root-Finding for Scalar Equations . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
582
583
587
587
589
589
591
592
593
594
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
598
598
598
600
602
603
605
606
610
612
613
613
616
617
617
619
620
621
623
625
628
629
630
631
633
B A Regularization Procedure
635
References
637
Index
653
This is page xvii
Printer: Opaque this
Preface
This is a book for people interested in solving optimization problems. Because of the wide
(and growing) use of optimization in science, engineering, economics, and industry, it is
essential for students and practitioners alike to develop an understanding of optimization
algorithms. Knowledge of the capabilities and limitations of these algorithms leads to a better
understanding of their impact on various applications, and points the way to future research
on improving and extending optimization algorithms and software. Our goal in this book
is to give a comprehensive description of the most powerful, state-of-the-art, techniques
for solving continuous optimization problems. By presenting the motivating ideas for each
algorithm, we try to stimulate the reader’s intuition and make the technical details easier to
follow. Formal mathematical requirements are kept to a minimum.
Because of our focus on continuous problems, we have omitted discussion of important optimization topics such as discrete and stochastic optimization. However, there are a
great many applications that can be formulated as continuous optimization problems; for
instance,
finding the optimal trajectory for an aircraft or a robot arm;
identifying the seismic properties of a piece of the earth’s crust by fitting a model of
the region under study to a set of readings from a network of recording stations;
xviii
PREFACE
designing a portfolio of investments to maximize expected return while maintaining
an acceptable level of risk;
controlling a chemical process or a mechanical device to optimize performance or
meet standards of robustness;
computing the optimal shape of an automobile or aircraft component.
Every year optimization algorithms are being called on to handle problems that
are much larger and complex than in the past. Accordingly, the book emphasizes largescale optimization techniques, such as interior-point methods, inexact Newton methods,
limited-memory methods, and the role of partially separable functions and automatic
differentiation. It treats important topics such as trust-region methods and sequential
quadratic programming more thoroughly than existing texts, and includes comprehensive
discussion of such “core curriculum” topics as constrained optimization theory, Newton
and quasi-Newton methods, nonlinear least squares and nonlinear equations, the simplex
method, and penalty and barrier methods for nonlinear programming.
The Audience
We intend that this book will be used in graduate-level courses in optimization, as offered in engineering, operations research, computer science, and mathematics departments.
There is enough material here for a two-semester (or three-quarter) sequence of courses.
We hope, too, that this book will be used by practitioners in engineering, basic science, and
industry, and our presentation style is intended to facilitate self-study. Since the book treats
a number of new algorithms and ideas that have not been described in earlier textbooks, we
hope that this book will also be a useful reference for optimization researchers.
Prerequisites for this book include some knowledge of linear algebra (including numerical linear algebra) and the standard sequence of calculus courses. To make the book as
self-contained as possible, we have summarized much of the relevant material from these areas in the Appendix. Our experience in teaching engineering students has shown us that the
material is best assimilated when combined with computer programming projects in which
the student gains a good feeling for the algorithms—their complexity, memory demands,
and elegance—and for the applications. In most chapters we provide simple computer
exercises that require only minimal programming proficiency.
Emphasis and Writing Style
We have used a conversational style to motivate the ideas and present the numerical
algorithms. Rather than being as concise as possible, our aim is to make the discussion flow
in a natural way. As a result, the book is comparatively long, but we believe that it can be
read relatively rapidly. The instructor can assign substantial reading assignments from the
text and focus in class only on the main ideas.
A typical chapter begins with a nonrigorous discussion of the topic at hand, including
figures and diagrams and excluding technical details as far as possible. In subsequent sections,
PREFACE
the algorithms are motivated and discussed, and then stated explicitly. The major theoretical
results are stated, and in many cases proved, in a rigorous fashion. These proofs can be
skipped by readers who wish to avoid technical details.
The practice of optimization depends not only on efficient and robust algorithms,
but also on good modeling techniques, careful interpretation of results, and user-friendly
software. In this book we discuss the various aspects of the optimization process—modeling,
optimality conditions, algorithms, implementation, and interpretation of results—but not
with equal weight. Examples throughout the book show how practical problems are formulated as optimization problems, but our treatment of modeling is light and serves mainly
to set the stage for algorithmic developments. We refer the reader to Dantzig [86] and
Fourer, Gay, and Kernighan [112] for more comprehensive discussion of this issue. Our
treatment of optimality conditions is thorough but not exhaustive; some concepts are discussed more extensively in Mangasarian [198] and Clarke [62]. As mentioned above, we are
quite comprehensive in discussing optimization algorithms.
Topics Not Covered
We omit some important topics, such as network optimization, integer programming,
stochastic programming, nonsmooth optimization, and global optimization. Network and
integer optimization are described in some excellent texts: for instance, Ahuja, Magnanti, and
Orlin [1] in the case of network optimization and Nemhauser and Wolsey [224], Papadimitriou and Steiglitz [235], and Wolsey [312] in the case of integer programming. Books on
stochastic optimization are only now appearing; we mention those of Kall and Wallace [174],
Birge and Louveaux [22]. Nonsmooth optimization comes in many flavors. The relatively
simple structures that arise in robust data fitting (which is sometimes based on the 1 norm)
are treated by Osborne [232] and Fletcher [101]. The latter book also discusses algorithms
for nonsmooth penalty functions that arise in constrained optimization; we discuss these
briefly, too, in Chapter 18. A more analytical treatment of nonsmooth optimization is given
by Hiriart-Urruty and Lemaréchal [170]. We omit detailed treatment of some important
topics that are the focus of intense current research, including interior-point methods for
nonlinear programming and algorithms for complementarity problems.
Additional Resource
The material in the book is complemented by an online resource called the NEOS
Guide, which can be found on the World-Wide Web at
http://www.mcs.anl.gov/otc/Guide/
The Guide contains information about most areas of optimization, and presents a number
of case studies that describe applications of various optimization algorithms to real-world
problems such as portfolio optimization and optimal dieting. Some of this material is
interactive in nature and has been used extensively for class exercises.
xix
xx
PREFACE
For the most part, we have omitted detailed discussions of specific software packages,
and refer the reader to Moré and Wright [217] or to the Software Guide section of the NEOS
Guide, which can be found at
http://www.mcs.anl.gov/otc/Guide/SoftwareGuide/
Users of optimization software refer in great numbers to this web site, which is being
constantly updated to reflect new packages and changes to existing software.
Acknowledgments
We are most grateful to the following colleagues for their input and feedback on various
sections of this work: Chris Bischof, Richard Byrd, George Corliss, Bob Fourer, David Gay,
Jean-Charles Gilbert, Phillip Gill, Jean-Pierre Goux, Don Goldfarb, Nick Gould, Andreas
Griewank, Matthias Heinkenschloss, Marcelo Marazzi, Hans Mittelmann, Jorge Moré, Will
Naylor, Michael Overton, Bob Plemmons, Hugo Scolnik, David Stewart, Philippe Toint,
Luis Vicente, Andreas Wächter, and Ya-xiang Yuan. We thank Guanghui Liu, who provided
help with many of the exercises, and Jill Lavelle who assisted us in preparing the figures. We
also express our gratitude to our sponsors at the Department of Energy and the National
Science Foundation, who have strongly supported our research efforts in optimization over
the years.
One of us (JN) would like to express his deep gratitude to Richard Byrd, who has taught
him so much about optimization and who has helped him in very many ways throughout
the course of his career.
Final Remark
In the preface to his 1987 book [101], Roger Fletcher described the field of optimization
as a “fascinating blend of theory and computation, heuristics and rigor.” The ever-growing
realm of applications and the explosion in computing power is driving optimization research
in new and exciting directions, and the ingredients identified by Fletcher will continue to
play important roles for many years to come.
Jorge Nocedal
Evanston, IL
Stephen J. Wright
Argonne, IL
This is page xxi
Printer: Opaque this
Preface to the
Second Edition
During the six years since the first edition of this book appeared, the field of continuous
optimization has continued to grow and evolve. This new edition reflects a better understanding of constrained optimization at both the algorithmic and theoretical levels, and of
the demands imposed by practical applications. Perhaps most notably, new chapters have
been added on two important topics: derivative-free optimization (Chapter 9) and interiorpoint methods for nonlinear programming (Chapter 19). The former topic has proved to
be of great interest in applications, while the latter topic has come into its own in recent
years and now forms the basis of successful codes for nonlinear programming.
Apart from the new chapters, we have revised and updated throughout the book,
de-emphasizing or omitting less important topics, enhancing the treatment of subjects of
evident interest, and adding new material in many places. The first part (unconstrained optimization) has been comprehensively reorganized to improve clarity. Discussion of Newton’s
method—the touchstone method for unconstrained problems—is distributed more naturally throughout this part rather than being isolated in a single chapter. An expanded
discussion of large-scale problems appears in Chapter 7.
Some reorganization has taken place also in the second part (constrained optimization), with material common to sequential quadratic programming and interior-point
methods now appearing in the chapter on fundamentals of nonlinear programming
xxii
PREFACE
TO THE
SECOND EDITION
algorithms (Chapter 15) and the discussion of primal barrier methods moved to the new
interior-point chapter. There is much new material in this part, including a treatment of
nonlinear programming duality, an expanded discussion of algorithms for inequality constrained quadratic programming, a discussion of dual simplex and presolving in linear
programming, a summary of practical issues in the implementation of interior-point linear
programming algorithms, a description of conjugate-gradient methods for quadratic programming, and a discussion of filter methods and nonsmooth penalty methods in nonlinear
programming algorithms.
In many chapters we have added a Perspectives and Software section near the end, to
place the preceding discussion in context and discuss the state of the art in software. The
appendix has been rearranged with some additional topics added, so that it can be used
in a more stand-alone fashion to cover some of the mathematical background required
for the rest of the book. The exercises have been revised in most chapters. After these
many additions, deletions, and changes, the second edition is only slightly longer than the
first, reflecting our belief that careful selection of the material to include and exclude is an
important responsibility for authors of books of this type.
A manual containing solutions for selected problems will be available to bona fide
instructors through the publisher. A list of typos will be maintained on the book’s web site,
which is accessible from the web pages of both authors.
We acknowledge with gratitude the comments and suggestions of many readers of the
first edition, who sent corrections to many errors and provided valuable perspectives on the
material, which led often to substantial changes. We mention in particular Frank Curtis,
Michael Ferris, Andreas Griewank, Jacek Gondzio, Sven Leyffer, Philip Loewen, Rembert
Reemtsen, and David Stewart.
Our special thanks goes to Michael Overton, who taught from a draft of the second
edition and sent many detailed and excellent suggestions. We also thank colleagues who
read various chapters of the new edition carefully during development, including Richard
Byrd, Nick Gould, Paul Hovland, Gabo Lopéz-Calva, Long Hei, Katya Scheinberg, Andreas
Wächter, and Richard Waltz. We thank Jill Wright for improving some of the figures and for
the new cover graphic.
We mentioned in the original preface several areas of optimization that are not
covered in this book. During the past six years, this list has only grown longer, as the field
has continued to expand in new directions. In this regard, the following areas are particularly
noteworthy: optimization problems with complementarity constraints, second-order cone
and semidefinite programming, simulation-based optimization, robust optimization, and
mixed-integer nonlinear programming. All these areas have seen theoretical and algorithmic
advances in recent years, and in many cases developments are being driven by new classes
of applications. Although this book does not cover any of these areas directly, it provides a
foundation from which they can be studied.
Jorge Nocedal
Evanston, IL
Stephen J. Wright
Madison, WI
This is page 1
Printer: Opaque this
CHAPTER
1
Introduction
People optimize. Investors seek to create portfolios that avoid excessive risk while achieving a
high rate of return. Manufacturers aim for maximum efficiency in the design and operation
of their production processes. Engineers adjust parameters to optimize the performance of
their designs.
Nature optimizes. Physical systems tend to a state of minimum energy. The molecules
in an isolated chemical system react with each other until the total potential energy of their
electrons is minimized. Rays of light follow paths that minimize their travel time.
2
CHAPTER 1.
INTRODUCTION
Optimization is an important tool in decision science and in the analysis of physical
systems. To make use of this tool, we must first identify some objective, a quantitative measure
of the performance of the system under study. This objective could be profit, time, potential
energy, or any quantity or combination of quantities that can be represented by a single
number. The objective depends on certain characteristics of the system, called variables or
unknowns. Our goal is to find values of the variables that optimize the objective. Often the
variables are restricted, or constrained, in some way. For instance, quantities such as electron
density in a molecule and the interest rate on a loan cannot be negative.
The process of identifying objective, variables, and constraints for a given problem is
known as modeling. Construction of an appropriate model is the first step—sometimes the
most important step—in the optimization process. If the model is too simplistic, it will not
give useful insights into the practical problem. If it is too complex, it may be too difficult to
solve.
Once the model has been formulated, an optimization algorithm can be used to
find its solution, usually with the help of a computer. There is no universal optimization
algorithm but rather a collection of algorithms, each of which is tailored to a particular type
of optimization problem. The responsibility of choosing the algorithm that is appropriate
for a specific application often falls on the user. This choice is an important one, as it may
determine whether the problem is solved rapidly or slowly and, indeed, whether the solution
is found at all.
After an optimization algorithm has been applied to the model, we must be able to
recognize whether it has succeeded in its task of finding a solution. In many cases, there
are elegant mathematical expressions known as optimality conditions for checking that the
current set of variables is indeed the solution of the problem. If the optimality conditions are
not satisfied, they may give useful information on how the current estimate of the solution
can be improved. The model may be improved by applying techniques such as sensitivity
analysis, which reveals the sensitivity of the solution to changes in the model and data.
Interpretation of the solution in terms of the application may also suggest ways in which the
model can be refined or improved (or corrected). If any changes are made to the model, the
optimization problem is solved anew, and the process repeats.
MATHEMATICAL FORMULATION
Mathematically speaking, optimization is the minimization or maximization of a
function subject to constraints on its variables. We use the following notation:
- x is the vector of variables, also called unknowns or parameters;
- f is the objective function, a (scalar) function of x that we want to maximize or
minimize;
- ci are constraint functions, which are scalar functions of x that define certain equations
and inequalities that the unknown vector x must satisfy.
CHAPTER 1.
INTRODUCTION
x2
c2
c1
contours of f
feasible
region
x*
x1
Figure 1.1
Geometrical representation of the problem (1.2).
Using this notation, the optimization problem can be written as follows:
minn f (x)
subject to
x∈IR
ci (x) 0,
i ∈ E,
ci (x) ≥ 0,
i ∈ I.
(1.1)
Here I and E are sets of indices for equality and inequality constraints, respectively.
As a simple example, consider the problem
min (x1 − 2)2 + (x2 − 1)2
subject to
x12 − x2 ≤ 0,
(1.2)
x1 + x2 ≤ 2.
We can write this problem in the form (1.1) by defining
f (x) (x1 − 2) + (x2 − 1) ,
2
c(x) c1 (x)
c2 (x)
x
2
−x12 + x2
−x1 − x2 + 2
x1
x2
,
,
I {1, 2},
E ∅.
Figure 1.1 shows the contours of the objective function, that is, the set of points for which
f (x) has a constant value. It also illustrates the feasible region, which is the set of points
satisfying all the constraints (the area between the two constraint boundaries), and the point
3
4
CHAPTER 1.
INTRODUCTION
x ∗ , which is the solution of the problem. Note that the “infeasible side” of the inequality
constraints is shaded.
The example above illustrates, too, that transformations are often necessary to express
an optimization problem in the particular form (1.1). Often it is more natural or convenient
to label the unknowns with two or three subscripts, or to refer to different variables by
completely different names, so that relabeling is necessary to pose the problem in the form
(1.1). Another common difference is that we are required to maximize rather than minimize
f , but we can accommodate this change easily by minimizing − f in the formulation (1.1).
Good modeling systems perform the conversion to standardized formulations such as (1.1)
transparently to the user.
EXAMPLE: A TRANSPORTATION PROBLEM
We begin with a much simplified example of a problem that might arise in manufacturing and transportation. A chemical company has 2 factories F1 and F2 and a dozen retail
outlets R1 , R2 , . . . , R12 . Each factory Fi can produce ai tons of a certain chemical product
each week; ai is called the capacity of the plant. Each retail outlet R j has a known weekly
demand of b j tons of the product. The cost of shipping one ton of the product from factory
Fi to retail outlet R j is ci j .
The problem is to determine how much of the product to ship from each factory
to each outlet so as to satisfy all the requirements and minimize cost. The variables of the
problem are xi j , i 1, 2, j 1, . . . , 12, where xi j is the number of tons of the product
shipped from factory Fi to retail outlet R j ; see Figure 1.2. We can write the problem as
min
(1.3a)
ci j xi j
ij
subject to
12
xi j ≤ ai ,
i 1, 2,
(1.3b)
xi j ≥ b j ,
j 1, . . . , 12,
(1.3c)
i 1, 2, j 1, . . . , 12.
(1.3d)
j1
2
i1
xi j ≥ 0,
This type of problem is known as a linear programming problem, since the objective function
and the constraints are all linear functions. In a more practical model, we would also include
costs associated with manufacturing and storing the product. There may be volume discounts
in practice for shipping the product; for example the cost (1.3a) could be represented by
i j ci j δ + x i j , where δ > 0 is a small subscription fee. In this case, the problem is a
nonlinear program because the objective function is nonlinear.
CHAPTER 1.
INTRODUCTION
R1
F1
X21
R2
R3
F2
R12
Figure 1.2
A transportation problem.
CONTINUOUS VERSUS DISCRETE OPTIMIZATION
In some optimization problems the variables make sense only if they take on integer
values. For example, a variable xi could represent the number of power plants of type i
that should be constructed by an electicity provider during the next 5 years, or it could
indicate whether or not a particular factory should be located in a particular city. The
mathematical formulation of such problems includes integrality constraints, which have
the form xi ∈ Z, where Z is the set of integers, or binary constraints, which have the form
xi ∈ {0, 1}, in addition to algebraic constraints like those appearing in (1.1). Problems of
this type are called integer programming problems. If some of the variables in the problem
are not restricted to be integer or binary variables, they are sometimes called mixed integer
programming problems, or MIPs for short.
Integer programming problems are a type of discrete optimization problem. Generally,
discrete optimization problems may contain not only integers and binary variables, but also
more abstract variable objects such as permutations of an ordered set. The defining feature
of a discrete optimization problem is that the unknown x is drawn from a a finite (but often
very large) set. By contrast, the feasible set for continuous optimization problems—the class
of problems studied in this book—is usually uncountably infinite, as when the components
of x are allowed to be real numbers. Continuous optimization problems are normally easier
to solve because the smoothness of the functions makes it possible to use objective and
constraint information at a particular point x to deduce information about the function’s
behavior at all points close to x. In discrete problems, by constrast, the behavior of the
objective and constraints may change significantly as we move from one feasible point to
another, even if the two points are “close” by some measure. The feasible sets for discrete
optimization problems can be thought of as exhibiting an extreme form of nonconvexity, as
a convex combination of two feasible points is in general not feasible.
5
6
CHAPTER 1.
INTRODUCTION
Discrete optimization problems are not addressed directly in this book; we refer the
reader to the texts by Papadimitriou and Steiglitz [235], Nemhauser and Wolsey [224], Cook
et al. [77], and Wolsey [312] for comprehensive treatments of this subject. We note, however,
that continuous optimization techniques often play an important role in solving discrete
optimization problems. For instance, the branch-and-bound method for integer linear
programming problems requires the repeated solution of linear programming “relaxations,”
in which some of the integer variables are fixed at integer values, while for other integer
variables the integrality constraints are temporarily ignored. These subproblems are usually
solved by the simplex method, which is discussed in Chapter 13 of this book.
CONSTRAINED AND UNCONSTRAINED OPTIMIZATION
Problems with the general form (1.1) can be classified according to the nature of the
objective function and constraints (linear, nonlinear, convex), the number of variables (large
or small), the smoothness of the functions (differentiable or nondifferentiable), and so on.
An important distinction is between problems that have constraints on the variables and
those that do not. This book is divided into two parts according to this classification.
Unconstrained optimization problems, for which we have E I ∅ in (1.1), arise
directly in many practical applications. Even for some problems with natural constraints
on the variables, it may be safe to disregard them as they do not affect on the solution and
do not interfere with algorithms. Unconstrained problems arise also as reformulations of
constrained optimization problems, in which the constraints are replaced by penalization
terms added to objective function that have the effect of discouraging constraint violations.
Constrained optimization problems arise from models in which constraints play an
essential role, for example in imposing budgetary constraints in an economic problem or
shape constraints in a design problem. These constraints may be simple bounds such as
0 ≤ x1 ≤ 100, more general linear constraints such as i xi ≤ 1, or nonlinear inequalities
that represent complex relationships among the variables.
When the objective function and all the constraints are linear functions of x, the
problem is a linear programming problem. Problems of this type are probably the most
widely formulated and solved of all optimization problems, particularly in management,
financial, and economic applications. Nonlinear programming problems, in which at least
some of the constraints or the objective are nonlinear functions, tend to arise naturally in
the physical sciences and engineering, and are becoming more widely used in management
and economic sciences as well.
GLOBAL AND LOCAL OPTIMIZATION
Many algorithms for nonlinear optimization problems seek only a local solution, a
point at which the objective function is smaller than at all other feasible nearby points. They
do not always find the global solution, which is the point with lowest function value among all
feasible points. Global solutions are needed in some applications, but for many problems they
CHAPTER 1.
INTRODUCTION
are difficult to recognize and even more difficult to locate. For convex programming problems,
and more particularly for linear programs, local solutions are also global solutions. General
nonlinear problems, both constrained and unconstrained, may possess local solutions that
are not global solutions.
In this book we treat global optimization only in passing and focus instead on the
computation and characterization of local solutions. We note, however, that many successful
global optimization algorithms require the solution of many local optimization problems,
to which the algorithms described in this book can be applied.
Research papers on global optimization can be found in Floudas and Pardalos [109]
and in the Journal of Global Optimization.
STOCHASTIC AND DETERMINISTIC OPTIMIZATION
In some optimization problems, the model cannot be fully specified because it depends
on quantities that are unknown at the time of formulation. This characteristic is shared by
many economic and financial planning models, which may depend for example on future
interest rates, future demands for a product, or future commodity prices, but uncertainty
can arise naturally in almost any type of application.
Rather than just use a “best guess” for the uncertain quantities, modelers may obtain
more useful solutions by incorporating additional knowledge about these quantities into
the model. For example, they may know a number of possible scenarios for the uncertain
demand, along with estimates of the probabilities of each scenario. Stochastic optimization
algorithms use these quantifications of the uncertainty to produce solutions that optimize
the expected performance of the model.
Related paradigms for dealing with uncertain data in the model include chanceconstrained optimization, in which we ensure that the variables x satisfy the given constraints
to some specified probability, and robust optimization, in which certain constraints are
required to hold for all possible values of the uncertain data.
We do not consider stochastic optimization problems further in this book, focusing
instead on deterministic optimization problems, in which the model is completely known.
Many algorithms for stochastic optimization do, however, proceed by formulating one or
more deterministic subproblems, each of which can be solved by the techniques outlined
here.
Stochastic and robust optimization have seen a great deal of recent research activity.
For further information on stochastic optimization, consult the books of Birge and
Louveaux [22] and Kall and Wallace [174]. Robust optimization is discussed in Ben-Tal
and Nemirovski [15].
CONVEXITY
The concept of convexity is fundamental in optimization. Many practical problems
possess this property, which generally makes them easier to solve both in theory and practice.
7
8
CHAPTER 1.
INTRODUCTION
The term “convex” can be applied both to sets and to functions. A set S ∈ IRn is a
convex set if the straight line segment connecting any two points in S lies entirely inside S.
Formally, for any two points x ∈ S and y ∈ S, we have αx + (1 − α)y ∈ S for all α ∈ [0, 1].
The function f is a convex function if its domain S is a convex set and if for any two points
x and y in S, the following property is satisfied:
f (αx + (1 − α)y) ≤ α f (x) + (1 − α) f (y),
for all α ∈ [0, 1].
(1.4)
Simple instances of convex sets include the unit ball {y ∈ IRn | y2 ≤ 1}; and any
polyhedron, which is a set defined by linear equalities and inequalities, that is,
{x ∈ IRn | Ax b, C x ≤ d},
where A and C are matrices of appropriate dimension, and b and d are vectors. Simple
instances of convex functions include the linear function f (x) c T x + α, for any constant
vector c ∈ IRn and scalar α; and the convex quadratic function f (x) x T H x, where H is
a symmetric positive semidefinite matrix.
We say that f is strictly convex if the inequality in (1.4) is strict whenever x y and
α is in the open interval (0, 1). A function f is said to be concave if − f is convex.
If the objective function in the optimization problem (1.1) and the feasible region are
both convex, then any local solution of the problem is in fact a global solution.
The term convex programming is used to describe a special case of the general
constrained optimization problem (1.1) in which
• the objective function is convex,
• the equality constraint functions ci (·), i ∈ E, are linear, and
• the inequality constraint functions ci (·), i ∈ I, are concave.
OPTIMIZATION ALGORITHMS
Optimization algorithms are iterative. They begin with an initial guess of the variable
x and generate a sequence of improved estimates (called “iterates”) until they terminate,
hopefully at a solution. The strategy used to move from one iterate to the next distinguishes
one algorithm from another. Most strategies make use of the values of the objective function
f , the constraint functions ci , and possibly the first and second derivatives of these functions.
Some algorithms accumulate information gathered at previous iterations, while others use
only local information obtained at the current point. Regardless of these specifics (which
will receive plenty of attention in the rest of the book), good algorithms should possess the
following properties:
• Robustness. They should perform well on a wide variety of problems in their class,
for all reasonable values of the starting point.
CHAPTER 1.
INTRODUCTION
• Efficiency. They should not require excessive computer time or storage.
• Accuracy. They should be able to identify a solution with precision, without being
overly sensitive to errors in the data or to the arithmetic rounding errors that occur
when the algorithm is implemented on a computer.
These goals may conflict. For example, a rapidly convergent method for a large unconstrained nonlinear problem may require too much computer storage. On the other hand,
a robust method may also be the slowest. Tradeoffs between convergence rate and storage
requirements, and between robustness and speed, and so on, are central issues in numerical
optimization. They receive careful consideration in this book.
The mathematical theory of optimization is used both to characterize optimal points
and to provide the basis for most algorithms. It is not possible to have a good understanding
of numerical optimization without a firm grasp of the supporting theory. Accordingly,
this book gives a solid (though not comprehensive) treatment of optimality conditions, as
well as convergence analysis that reveals the strengths and weaknesses of some of the most
important algorithms.
NOTES AND REFERENCES
Optimization traces its roots to the calculus of variations and the work of Euler and
Lagrange. The development of linear programming n the 1940s broadened the field and
stimulated much of the progress in modern optimization theory and practice during the
past 60 years.
Optimization is often called mathematical programming, a somewhat confusing term
coined in the 1940s, before the word “programming” became inextricably linked with
computer software. The original meaning of this word (and the intended one in this context)
was more inclusive, with connotations of algorithm design and analysis.
Modeling will not be treated extensively in the book. It is an essential subject in its
own right, as it makes the connection between optimization algorithms and software on
the one hand, and applications on the other hand. Information about modeling techniques
for various application areas can be found in Dantzig [86], Ahuja, Magnanti, and Orlin [1],
Fourer, Gay, and Kernighan [112], Winston [308], and Rardin [262].
9
This is pag
Printer: O
CHAPTER
2
Fundamentals of
Unconstrained
Optimization
In unconstrained optimization, we minimize an objective function that depends on real
variables, with no restrictions at all on the values of these variables. The mathematical
formulation is
min f (x),
x
(2.1)
where x ∈ IRn is a real vector with n ≥ 1 components and f : IRn → IR is a smooth
function.
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
y
.
y3
y2
y1
.
. .
.
t
t1
Figure 2.1
t2
t3
tm
Least squares data fitting problem.
Usually, we lack a global perspective on the function f . All we know are the values
of f and maybe some of its derivatives at a set of points x0 , x1 , x2 , . . .. Fortunately, our
algorithms get to choose these points, and they try to do so in a way that identifies a solution
reliably and without using too much computer time or storage. Often, the information
about f does not come cheaply, so we usually prefer algorithms that do not call for this
information unnecessarily.
❏ EXAMPLE 2.1
Suppose that we are trying to find a curve that fits some experimental data. Figure 2.1
plots measurements y1 , y2 , . . . , ym of a signal taken at times t1 , t2 , . . . , tm . From the data and
our knowledge of the application, we deduce that the signal has exponential and oscillatory
behavior of certain types, and we choose to model it by the function
φ(t; x) x1 + x2 e−(x3 −t) /x4 + x5 cos(x6 t).
2
The real numbers xi , i 1, 2, . . . , 6, are the parameters of the model; we would like to
choose them to make the model values φ(t j ; x) fit the observed data y j as closely as possible.
To state our objective as an optimization problem, we group the parameters xi into a vector
of unknowns x (x1 , x2 , . . . , x6 )T , and define the residuals
r j (x) y j − φ(t j ; x),
j 1, 2, . . . , m,
(2.2)
which measure the discrepancy between the model and the observed data. Our estimate of
11
12
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
x will be obtained by solving the problem
min6 f (x) r12 (x) + r22 (x) + · · · + rm2 (x).
(2.3)
x∈IR
This is a nonlinear least-squares problem, a special case of unconstrained optimization.
It illustrates that some objective functions can be expensive to evaluate even when the
number of variables is small. Here we have n 6, but if the number of measurements
m is large (105 , say), evaluation of f (x) for a given parameter vector x is a significant
computation.
❐
Suppose that for the data given in Figure 2.1 the optimal solution of (2.3) is approximately x ∗ (1.1, 0.01, 1.2, 1.5, 2.0, 1.5) and the corresponding function value is
f (x ∗ ) 0.34. Because the optimal objective is nonzero, there must be discrepancies between the observed measurements y j and the model predictions φ(t j , x ∗ ) for some (usually
most) values of j—the model has not reproduced all the data points exactly. How, then,
can we verify that x ∗ is indeed a minimizer of f ? To answer this question, we need to
define the term “solution” and explain how to recognize solutions. Only then can we discuss
algorithms for unconstrained optimization problems.
2.1
WHAT IS A SOLUTION?
Generally, we would be happiest if we found a global minimizer of f , a point where the
function attains its least value. A formal definition is
A point x ∗ is a global minimizer if f (x ∗ ) ≤ f (x) for all x,
where x ranges over all of IRn (or at least over the domain of interest to the modeler). The
global minimizer can be difficult to find, because our knowledge of f is usually only local.
Since our algorithm does not visit many points (we hope!), we usually do not have a good
picture of the overall shape of f , and we can never be sure that the function does not take a
sharp dip in some region that has not been sampled by the algorithm. Most algorithms are
able to find only a local minimizer, which is a point that achieves the smallest value of f in
its neighborhood. Formally, we say:
A point x ∗ is a local minimizer if there is a neighborhood N of x ∗ such that f (x ∗ ) ≤
f (x) for all x ∈ N .
(Recall that a neighborhood of x ∗ is simply an open set that contains x ∗ .) A point that satisfies
this definition is sometimes called a weak local minimizer. This terminology distinguishes
2.1.
WHAT IS
A
SOLUTION?
it from a strict local minimizer, which is the outright winner in its neighborhood.
Formally,
A point x ∗ is a strict local minimizer (also called a strong local minimizer) if there is a
neighborhood N of x ∗ such that f (x ∗ ) < f (x) for all x ∈ N with x x ∗ .
For the constant function f (x) 2, every point x is a weak local minimizer, while the
function f (x) (x − 2)4 has a strict local minimizer at x 2.
A slightly more exotic type of local minimizer is defined as follows.
A point x ∗ is an isolated local minimizer if there is a neighborhood N of x ∗ such that
x ∗ is the only local minimizer in N .
Some strict local minimizers are not isolated, as illustrated by the function
f (x) x 4 cos(1/x) + 2x 4 ,
f (0) 0,
which is twice continuously differentiable and has a strict local minimizer at x ∗ 0.
However, there are strict local minimizers at many nearby points x j , and we can label these
points so that x j → 0 as j → ∞.
While strict local minimizers are not always isolated, it is true that all isolated local
minimizers are strict.
Figure 2.2 illustrates a function with many local minimizers. It is usually difficult
to find the global minimizer for such functions, because algorithms tend to be “trapped”
at local minimizers. This example is by no means pathological. In optimization problems
associated with the determination of molecular conformation, the potential function to be
minimized may have millions of local minima.
f
x
Figure 2.2
A difficult case for global minimization.
13
14
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
Sometimes we have additional “global” knowledge about f that may help in identifying global minima. An important special case is that of convex functions, for which every
local minimizer is also a global minimizer.
RECOGNIZING A LOCAL MINIMUM
From the definitions given above, it might seem that the only way to find out whether
a point x ∗ is a local minimum is to examine all the points in its immediate vicinity, to
make sure that none of them has a smaller function value. When the function f is smooth,
however, there are more efficient and practical ways to identify local minima. In particular, if
f is twice continuously differentiable, we may be able to tell that x ∗ is a local minimizer (and
possibly a strict local minimizer) by examining just the gradient ∇ f (x ∗ ) and the Hessian
∇ 2 f (x ∗ ).
The mathematical tool used to study minimizers of smooth functions is Taylor’s
theorem. Because this theorem is central to our analysis throughout the book, we state it
now. Its proof can be found in any calculus textbook.
Theorem 2.1 (Taylor’s Theorem).
Suppose that f : IRn → IR is continuously differentiable and that p ∈ IRn . Then we have
that
f (x + p) f (x) + ∇ f (x + t p)T p,
(2.4)
for some t ∈ (0, 1). Moreover, if f is twice continuously differentiable, we have that
∇ f (x + p) ∇ f (x) +
1
∇ 2 f (x + t p) p dt,
(2.5)
0
and that
f (x + p) f (x) + ∇ f (x)T p + 12 p T ∇ 2 f (x + t p) p,
(2.6)
for some t ∈ (0, 1).
Necessary conditions for optimality are derived by assuming that x ∗ is a local minimizer
and then proving facts about ∇ f (x ∗ ) and ∇ 2 f (x ∗ ).
Theorem 2.2 (First-Order Necessary Conditions).
If x ∗ is a local minimizer and f is continuously differentiable in an open neighborhood
∗
of x , then ∇ f (x ∗ ) 0.
2.1.
WHAT IS
A
SOLUTION?
PROOF. Suppose for contradiction that ∇ f (x ∗ ) 0. Define the vector p −∇ f (x ∗ ) and
note that p T ∇ f (x ∗ ) −∇ f (x ∗ )2 < 0. Because ∇ f is continuous near x ∗ , there is a
scalar T > 0 such that
p T ∇ f (x ∗ + t p) < 0,
for all t ∈ [0, T ].
For any t¯ ∈ (0, T ], we have by Taylor’s theorem that
f (x ∗ + t¯ p) f (x ∗ ) + t¯ p T ∇ f (x ∗ + t p),
for some t ∈ (0, t¯).
Therefore, f (x ∗ + t¯ p) < f (x ∗ ) for all t¯ ∈ (0, T ]. We have found a direction leading
away from x ∗ along which f decreases, so x ∗ is not a local minimizer, and we have a
contradiction.
We call x ∗ a stationary point if ∇ f (x ∗ ) 0. According to Theorem 2.2, any local
minimizer must be a stationary point.
For the next result we recall that a matrix B is positive definite if p T Bp > 0 for all
p 0, and positive semidefinite if p T Bp ≥ 0 for all p (see the Appendix).
Theorem 2.3 (Second-Order Necessary Conditions).
If x ∗ is a local minimizer of f and ∇ 2 f exists and is continuous in an open neighborhood
∗
of x , then ∇ f (x ∗ ) 0 and ∇ 2 f (x ∗ ) is positive semidefinite.
PROOF. We know from Theorem 2.2 that ∇ f (x ∗ ) 0. For contradiction, assume
that ∇ 2 f (x ∗ ) is not positive semidefinite. Then we can choose a vector p such that
p T ∇ 2 f (x ∗ ) p < 0, and because ∇ 2 f is continuous near x ∗ , there is a scalar T > 0
such that p T ∇ 2 f (x ∗ + t p) p < 0 for all t ∈ [0, T ].
By doing a Taylor series expansion around x ∗ , we have for all t¯ ∈ (0, T ] and some
t ∈ (0, t¯) that
f (x ∗ + t¯ p) f (x ∗ ) + t¯ p T ∇ f (x ∗ ) + 12 t¯2 p T ∇ 2 f (x ∗ + t p) p < f (x ∗ ).
As in Theorem 2.2, we have found a direction from x ∗ along which f is decreasing, and so
again, x ∗ is not a local minimizer.
We now describe sufficient conditions, which are conditions on the derivatives of f at
the point z ∗ that guarantee that x ∗ is a local minimizer.
15
16
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
Theorem 2.4 (Second-Order Sufficient Conditions).
Suppose that ∇ 2 f is continuous in an open neighborhood of x ∗ and that ∇ f (x ∗ ) 0
and ∇ 2 f (x ∗ ) is positive definite. Then x ∗ is a strict local minimizer of f .
PROOF. Because the Hessian is continuous and positive definite at x ∗ , we can choose a radius
r > 0 so that ∇ 2 f (x) remains positive definite for all x in the open ball D {z | z − x ∗ <
r }. Taking any nonzero vector p with p < r , we have x ∗ + p ∈ D and so
f (x ∗ + p) f (x ∗ ) + p T ∇ f (x ∗ ) + 12 p T ∇ 2 f (z) p
f (x ∗ ) + 12 p T ∇ 2 f (z) p,
where z x ∗ + t p for some t ∈ (0, 1). Since z ∈ D, we have p T ∇ 2 f (z) p > 0, and therefore
f (x ∗ + p) > f (x ∗ ), giving the result.
Note that the second-order sufficient conditions of Theorem 2.4 guarantee something
stronger than the necessary conditions discussed earlier; namely, that the minimizer is a strict
local minimizer. Note too that the second-order sufficient conditions are not necessary: A
point x ∗ may be a strict local minimizer, and yet may fail to satisfy the sufficient conditions.
A simple example is given by the function f (x) x 4 , for which the point x ∗ 0 is a
strict local minimizer at which the Hessian matrix vanishes (and is therefore not positive
definite).
When the objective function is convex, local and global minimizers are simple to
characterize.
Theorem 2.5.
When f is convex, any local minimizer x ∗ is a global minimizer of f . If in addition f is
differentiable, then any stationary point x ∗ is a global minimizer of f .
PROOF. Suppose that x ∗ is a local but not a global minimizer. Then we can find a point
z ∈ IRn with f (z) < f (x ∗ ). Consider the line segment that joins x ∗ to z, that is,
x λz + (1 − λ)x ∗ ,
for some λ ∈ (0, 1].
(2.7)
By the convexity property for f , we have
f (x) ≤ λ f (z) + (1 − λ) f (x ∗ ) < f (x ∗ ).
(2.8)
Any neighborhood N of x ∗ contains a piece of the line segment (2.7), so there will always
be points x ∈ N at which (2.8) is satisfied. Hence, x ∗ is not a local minimizer.
2.1.
WHAT IS
A
SOLUTION?
For the second part of the theorem, suppose that x ∗ is not a global minimizer and
choose z as above. Then, from convexity, we have
d
f (x ∗ + λ(z − x ∗ )) |λ0 (see the Appendix)
dλ
f (x ∗ + λ(z − x ∗ )) − f (x ∗ )
lim
λ↓0
λ
λ f (z) + (1 − λ) f (x ∗ ) − f (x ∗ )
≤ lim
λ↓0
λ
f (z) − f (x ∗ ) < 0.
∇ f (x ∗ )T (z − x ∗ ) Therefore, ∇ f (x ∗ ) 0, and so x ∗ is not a stationary point.
These results, which are based on elementary calculus, provide the foundations for
unconstrained optimization algorithms. In one way or another, all algorithms seek a point
where ∇ f (·) vanishes.
NONSMOOTH PROBLEMS
This book focuses on smooth functions, by which we generally mean functions whose
second derivatives exist and are continuous. We note, however, that there are interesting
problems in which the functions involved may be nonsmooth and even discontinuous. It is
not possible in general to identify a minimizer of a general discontinuous function. If, however, the function consists of a few smooth pieces, with discontinuities between the pieces,
it may be possible to find the minimizer by minimizing each smooth piece individually.
If the function is continuous everywhere but nondifferentiable at certain points,
as in Figure 2.3, we can identify a solution by examing the subgradient or generalized
f
x*
Figure 2.3
x
Nonsmooth function with minimum at a kink.
17
18
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
gradient, which are generalizations of the concept of gradient to the nonsmooth case.
Nonsmooth optimization is beyond the scope of this book; we refer instead to HiriartUrruty and Lemaréchal [170] for an extensive discussion of theory. Here, we mention
only that the minimization of a function such as the one illustrated in Figure 2.3 (which
contains a jump discontinuity in the first derivative f (x) at the minimum) is difficult
because the behavior of f is not predictable near the point of nonsmoothness. That
is, we cannot be sure that information about f obtained at one point can be used
to infer anything about f at neighboring points, because points of nondifferentiability may intervene. However, minimization of certain special nondifferentiable functions,
such as
f (x) r (x)1 ,
f (x) r (x)∞
(2.9)
(where r (x) is a vector function), can be reformulated as smooth constrained optimization problems; see Exercise 12.5 in Chapter 12 and (17.31). The functions (2.9) are
useful in data fitting, where r (x) is the residual vector whose components are defined
in (2.2).
2.2
OVERVIEW OF ALGORITHMS
The last forty years have seen the development of a powerful collection of algorithms for
unconstrained optimization of smooth functions. We now give a broad description of their
main properties, and we describe them in more detail in Chapters 3, 4, 5, 6, and 7. All
algorithms for unconstrained minimization require the user to supply a starting point,
which we usually denote by x0 . The user with knowledge about the application and the
data set may be in a good position to choose x0 to be a reasonable estimate of the solution.
Otherwise, the starting point must be chosen by the algorithm, either by a systematic
approach or in some arbitrary manner.
Beginning at x0 , optimization algorithms generate a sequence of iterates {xk }∞
k0
that terminate when either no more progress can be made or when it seems that a solution point has been approximated with sufficient accuracy. In deciding how to move
from one iterate xk to the next, the algorithms use information about the function f at
xk , and possibly also information from earlier iterates x0 , x1 , . . . , xk−1 . They use this information to find a new iterate xk+1 with a lower function value than xk . (There exist
nonmonotone algorithms that do not insist on a decrease in f at every step, but even these
algorithms require f to be decreased after some prescribed number m of iterations, that is,
f (xk ) < f (xk−m ).)
There are two fundamental strategies for moving from the current point xk to a new
iterate xk+1 . Most of the algorithms described in this book follow one of these approaches.
2.2.
OVERVIEW
OF
ALGORITHMS
TWO STRATEGIES: LINE SEARCH AND TRUST REGION
In the line search strategy, the algorithm chooses a direction pk and searches along
this direction from the current iterate xk for a new iterate with a lower function value.
The distance to move along pk can be found by approximately solving the following onedimensional minimization problem to find a step length α:
min f (xk + αpk ).
(2.10)
α>0
By solving (2.10) exactly, we would derive the maximum benefit from the direction pk , but
an exact minimization may be expensive and is usually unnecessary. Instead, the line search
algorithm generates a limited number of trial step lengths until it finds one that loosely
approximates the minimum of (2.10). At the new point, a new search direction and step
length are computed, and the process is repeated.
In the second algorithmic strategy, known as trust region, the information gathered
about f is used to construct a model function m k whose behavior near the current point
xk is similar to that of the actual objective function f . Because the model m k may not be a
good approximation of f when x is far from xk , we restrict the search for a minimizer of m k
to some region around xk . In other words, we find the candidate step p by approximately
solving the following subproblem:
min m k (xk + p),
where xk + p lies inside the trust region.
p
(2.11)
If the candidate solution does not produce a sufficient decrease in f , we conclude that the
trust region is too large, and we shrink it and re-solve (2.11). Usually, the trust region is a
ball defined by p2 ≤ , where the scalar > 0 is called the trust-region radius. Elliptical
and box-shaped trust regions may also be used.
The model m k in (2.11) is usually defined to be a quadratic function of the form
m k (xk + p) f k + p T ∇ f k + 12 p T Bk p,
(2.12)
where f k , ∇ f k , and Bk are a scalar, vector, and matrix, respectively. As the notation indicates,
f k and ∇ f k are chosen to be the function and gradient values at the point xk , so that m k
and f are in agreement to first order at the current iterate xk . The matrix Bk is either the
Hessian ∇ 2 f k or some approximation to it.
Suppose that the objective function is given by f (x) 10(x2 − x12 )2 + (1 − x1 )2 . At
the point xk (0, 1) its gradient and Hessian are
∇ fk −2
20
,
∇ fk 2
−38
0
0
20
.
19
20
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
m = 12
m =1
contours
of f
xk
p
k
contours
of model
pk
*
unconstrained
minimizer
Figure 2.4 Two possible trust regions (circles) and their corresponding steps pk . The
solid lines are contours of the model function m k .
The contour lines of the quadratic model (2.12) with Bk ∇ 2 f k are depicted in Figure 2.4,
which also illustrates the contours of the objective function f and the trust region. We
have indicated contour lines where the model m k has values 1 and 12. Note from Figure 2.4
that each time we decrease the size of the trust region after failure of a candidate iterate,
the step from xk to the new candidate will be shorter, and it usually points in a different
direction from the previous candidate. The trust-region strategy differs in this respect from
line search, which stays with a single search direction.
In a sense, the line search and trust-region approaches differ in the order in which they
choose the direction and distance of the move to the next iterate. Line search starts by fixing
the direction pk and then identifying an appropriate distance, namely the step length αk . In
trust region, we first choose a maximum distance—the trust-region radius k —and then
seek a direction and step that attain the best improvement possible subject to this distance
constraint. If this step proves to be unsatisfactory, we reduce the distance measure k and
try again.
The line search approach is discussed in more detail in Chapter 3. Chapter 4 discusses
the trust-region strategy, including techniques for choosing and adjusting the size of the region and for computing approximate solutions to the trust-region problems (2.11). We now
preview two major issues: choice of the search direction pk in line search methods, and choice
of the Hessian Bk in trust-region methods. These issues are closely related, as we now observe.
SEARCH DIRECTIONS FOR LINE SEARCH METHODS
The steepest descent direction −∇ f k is the most obvious choice for search direction
for a line search method. It is intuitive; among all the directions we could move from xk ,
2.2.
OVERVIEW
OF
ALGORITHMS
it is the one along which f decreases most rapidly. To verify this claim, we appeal again
to Taylor’s theorem (Theorem 2.1), which tells us that for any search direction p and
step-length parameter α, we have
f (xk + αp) f (xk ) + αp T ∇ f k + 12 α 2 p T ∇ 2 f (xk + t p) p,
for some t ∈ (0, α)
(see (2.6)). The rate of change in f along the direction p at xk is simply the coefficient of
α, namely, p T ∇ f k . Hence, the unit direction p of most rapid decrease is the solution to the
problem
min p T ∇ f k ,
subject to p 1.
p
(2.13)
Since p T ∇ f k p ∇ f k cos θ ∇ f k cos θ , where θ is the angle between p and ∇ f k ,
it is easy to see that the minimizer is attained when cos θ −1 and
p −∇ f k /∇ f k ,
as claimed. As we illustrate in Figure 2.5, this direction is orthogonal to the contours of the
function.
The steepest descent method is a line search method that moves along pk −∇ f k at
every step. It can choose the step length αk in a variety of ways, as we discuss in Chapter 3. One
advantage of the steepest descent direction is that it requires calculation of the gradient ∇ f k
but not of second derivatives. However, it can be excruciatingly slow on difficult problems.
Line search methods may use search directions other than the steepest descent direction. In general, any descent direction—one that makes an angle of strictly less than π/2
radians with −∇ f k —is guaranteed to produce a decrease in f , provided that the step length
.
pk
x*
xk
Figure 2.5
Steepest descent direction for a function of two variables.
21
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
p
k
–
∆
22
fk
Figure 2.6
A downhill direction pk .
is sufficiently small (see Figure 2.6). We can verify this claim by using Taylor’s theorem.
From (2.6), we have that
f (xk + pk ) f (xk ) + pkT ∇ f k + O( 2 ).
When pk is a downhill direction, the angle θk between pk and ∇ f k has cos θk < 0, so that
pkT ∇ f k pk ∇ f k cos θk < 0.
It follows that f (xk + pk ) < f (xk ) for all positive but sufficiently small values of .
Another important search direction—perhaps the most important one of all—
is the Newton direction. This direction is derived from the second-order Taylor series
approximation to f (xk + p), which is
def
f (xk + p) ≈ f k + p T ∇ f k + 12 p T ∇ 2 f k p m k ( p).
(2.14)
Assuming for the moment that ∇ 2 f k is positive definite, we obtain the Newton direction
by finding the vector p that minimizes m k ( p). By simply setting the derivative of m k ( p) to
zero, we obtain the following explicit formula:
−1
pkN − ∇ 2 f k
∇ fk .
(2.15)
The Newton direction is reliable when the difference between the true function
f (xk + p) and its quadratic model m k ( p) is not too large. By comparing (2.14) with (2.6),
we see that the only difference between these functions is that the matrix ∇ 2 f (xk + t p) in
the third term of the expansion has been replaced by ∇ 2 f k . If ∇ 2 f is sufficiently smooth,
this difference introduces a perturbation of only O( p3 ) into the expansion, so that when
p is small, the approximation f (xk + p) ≈ m k ( p) is quite accurate.
2.2.
OVERVIEW
OF
ALGORITHMS
The Newton direction can be used in a line search method when ∇ 2 f k is positive
definite, for in this case we have
∇ f kT pkN − pkN T ∇ 2 f k pkN ≤ −σk pkN 2
for some σk > 0. Unless the gradient ∇ f k (and therefore the step pkN ) is zero, we have that
∇ f kT pkN < 0, so the Newton direction is a descent direction.
Unlike the steepest descent direction, there is a “natural” step length of 1 associated
with the Newton direction. Most line search implementations of Newton’s method use the
unit step α 1 where possible and adjust α only when it does not produce a satisfactory
reduction in the value of f .
When ∇ 2 f is not positive definite, the Newton direction may not even be defined,
2 −1 k
may not exist. Even when it is defined, it may not satisfy the descent property
since ∇ f k
∇ f kT pkN < 0, in which case it is unsuitable as a search direction. In these situations, line
search methods modify the definition of pk to make it satisfy the descent condition while
retaining the benefit of the second-order information contained in ∇ 2 f k . We describe these
modifications in Chapter 3.
Methods that use the Newton direction have a fast rate of local convergence, typically
quadratic. After a neighborhood of the solution is reached, convergence to high accuracy
often occurs in just a few iterations. The main drawback of the Newton direction is the
need for the Hessian ∇ 2 f (x). Explicit computation of this matrix of second derivatives
can sometimes be a cumbersome, error-prone, and expensive process. Finite-difference and
automatic differentiation techniques described in Chapter 8 may be useful in avoiding the
need to calculate second derivatives by hand.
Quasi-Newton search directions provide an attractive alternative to Newton’s method
in that they do not require computation of the Hessian and yet still attain a superlinear rate
of convergence. In place of the true Hessian ∇ 2 f k , they use an approximation Bk , which is
updated after each step to take account of the additional knowledge gained during the step.
The updates make use of the fact that changes in the gradient g provide information about
the second derivative of f along the search direction. By using the expression (2.5) from our
statement of Taylor’s theorem, we have by adding and subtracting the term ∇ 2 f (x) p that
∇ f (x + p) ∇ f (x) + ∇ f (x) p +
2
1
∇ 2 f (x + t p) − ∇ 2 f (x) p dt.
0
Because ∇ f (·) is continuous, the size of the final integral term is o( p). By setting x xk
and p xk+1 − xk , we obtain
∇ f k+1 ∇ f k + ∇ 2 f k (xk+1 − xk ) + o(xk+1 − xk ).
When xk and xk+1 lie in a region near the solution x ∗ , within which ∇ 2 f is positive definite,
the final term in this expansion is eventually dominated by the ∇ 2 f k (xk+1 − xk ) term, and
23
24
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
we can write
∇ 2 f k (xk+1 − xk ) ≈ ∇ f k+1 − ∇ f k .
(2.16)
We choose the new Hessian approximation Bk+1 so that it mimics the property (2.16) of
the true Hessian, that is, we require it to satisfy the following condition, known as the secant
equation:
Bk+1 sk yk ,
(2.17)
where
sk xk+1 − xk ,
yk ∇ f k+1 − ∇ f k .
Typically, we impose additional conditions on Bk+1 , such as symmetry (motivated by
symmetry of the exact Hessian), and a requirement that the difference between successive
approximations Bk and Bk+1 have low rank.
Two of the most popular formulae for updating the Hessian approximation Bk are
the symmetric-rank-one (SR1) formula, defined by
Bk+1 Bk +
(yk − Bk sk )(yk − Bk sk )T
,
(yk − Bk sk )T sk
(2.18)
and the BFGS formula, named after its inventors, Broyden, Fletcher, Goldfarb, and Shanno,
which is defined by
Bk+1 Bk −
Bk sk skT Bk
yk y T
+ Tk .
T
sk Bk sk
yk sk
(2.19)
Note that the difference between the matrices Bk and Bk+1 is a rank-one matrix in the
case of (2.18) and a rank-two matrix in the case of (2.19). Both updates satisfy the secant
equation and both maintain symmetry. One can show that BFGS update (2.19) generates
positive definite approximations whenever the initial approximation B0 is positive definite
and skT yk > 0. We discuss these issues further in Chapter 6.
The quasi-Newton search direction is obtained by using Bk in place of the exact
Hessian in the formula (2.15), that is,
pk −Bk−1 ∇ f k .
(2.20)
Some practical implementations of quasi-Newton methods avoid the need to factorize Bk
at each iteration by updating the inverse of Bk , instead of Bk itself. In fact, the equivalent
2.2.
OVERVIEW
OF
ALGORITHMS
formula for (2.18) and (2.19), applied to the inverse approximation Hk Bk−1 , is
def
Hk+1 I − ρk sk ykT Hk I − ρk yk skT + ρk sk skT ,
ρk 1
ykT sk
.
(2.21)
Calculation of pk can then be performed by using the formula pk −Hk ∇ f k . This matrix–
vector multiplication is simpler than the factorization/back-substitution procedure that is
needed to implement the formula (2.20).
Two variants of quasi-Newton methods designed to solve large problems—partially
separable and limited-memory updating—are described in Chapter 7.
The last class of search directions we preview here is that generated by nonlinear
conjugate gradient methods. They have the form
pk −∇ f (xk ) + βk pk−1 ,
where βk is a scalar that ensures that pk and pk−1 are conjugate—an important concept
in the minimization of quadratic functions that will be defined in Chapter 5. Conjugate
gradient methods were originally designed to solve systems of linear equations Ax b,
where the coefficient matrix A is symmetric and positive definite. The problem of solving
this linear system is equivalent to the problem of minimizing the convex quadratic function
defined by
φ(x) 12 x T Ax − b T x,
so it was natural to investigate extensions of these algorithms to more general types of
unconstrained minimization problems. In general, nonlinear conjugate gradient directions
are much more effective than the steepest descent direction and are almost as simple to
compute. These methods do not attain the fast convergence rates of Newton or quasiNewton methods, but they have the advantage of not requiring storage of matrices. An
extensive discussion of nonlinear conjugate gradient methods is given in Chapter 5.
All of the search directions discussed so far can be used directly in a line search
framework. They give rise to the steepest descent, Newton, quasi-Newton, and conjugate
gradient line search methods. All except conjugate gradients have an analogue in the trustregion framework, as we now discuss.
MODELS FOR TRUST-REGION METHODS
If we set Bk 0 in (2.12) and define the trust region using the Euclidean norm, the
trust-region subproblem (2.11) becomes
min f k + p T ∇ f k
p
subject to p2 ≤ k .
25
26
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
We can write the solution to this problem in closed form as
pk −
k ∇ f k
.
∇ f k This is simply a steepest descent step in which the step length is determined by the trustregion radius; the trust-region and line search approaches are essentially the same in this case.
A more interesting trust-region algorithm is obtained by choosing Bk to be the
exact Hessian ∇ 2 f k in the quadratic model (2.12). Because of the trust-region restriction
p2 ≤ k , the subproblem (2.11) is guaranteed to have a solution even when ∇ 2 f k is not
positive definite pk , as we see in Figure 2.4. The trust-region Newton method has proved to
be highly effective in practice, as we discuss in Chapter 7.
If the matrix Bk in the quadratic model function m k of (2.12) is defined by means of
a quasi-Newton approximation, we obtain a trust-region quasi-Newton method.
SCALING
The performance of an algorithm may depend crucially on how the problem is formulated. One important issue in problem formulation is scaling. In unconstrained optimization,
a problem is said to be poorly scaled if changes to x in a certain direction produce much larger
variations in the value of f than do changes to x in another direction. A simple example is
provided by the function f (x) 109 x12 + x22 , which is very sensitive to small changes in x1
but not so sensitive to perturbations in x2 .
Poorly scaled functions arise, for example, in simulations of physical and chemical
systems where different processes are taking place at very different rates. To be more specific,
consider a chemical system in which four reactions occur. Associated with each reaction is
a rate constant that describes the speed at which the reaction takes place. The optimization
problem is to find values for these rate constants by observing the concentrations of each
chemical in the system at different times. The four constants differ greatly in magnitude, since
the reactions take place at vastly different speeds. Suppose we have the following rough estimates for the final values of the constants, each correct to within, say, an order of magnitude:
x1 ≈ 10−10 ,
x2 ≈ x3 ≈ 1,
x4 ≈ 105 .
Before solving this problem we could introduce a new variable z defined by
⎡
x1
⎢
⎢ x2
⎢
⎢ x
⎣ 3
x4
⎤
⎡
⎥ ⎢
⎥ ⎢
⎥⎢
⎥ ⎢
⎦ ⎣
10−10
0
0
0
1
0
0
0
1
0
0
0
0
⎤⎡
z1
⎥⎢
⎥ ⎢ z2
⎥⎢
⎢
0 ⎥
⎦ ⎣ z3
105
z4
0
⎤
⎥
⎥
⎥,
⎥
⎦
and then define and solve the optimization problem in terms of the new variable z. The
2.2.
OVERVIEW
OF
ALGORITHMS
∆
– f
k
∆
– f
k
Figure 2.7 Poorly scaled and well scaled problems, and performance of the steepest
descent direction.
optimal values of z will be within about an order of magnitude of 1, making the solution
more balanced. This kind of scaling of the variables is known as diagonal scaling.
Scaling is performed (sometimes unintentionally) when the units used to represent
variables are changed. During the modeling process, we may decide to change the units of
some variables, say from meters to millimeters. If we do, the range of those variables and
their size relative to the other variables will both change.
Some optimization algorithms, such as steepest descent, are sensitive to poor scaling,
while others, such as Newton’s method, are unaffected by it. Figure 2.7 shows the contours
of two convex nearly quadratic functions, the first of which is poorly scaled, while the second
is well scaled. For the poorly scaled problem, the one with highly elongated contours, the
steepest descent direction does not yield much reduction in the function, while for the
well-scaled problem it performs much better. In both cases, Newton’s method will produce
a much better step, since the second-order quadratic model (m k in (2.14)) happens to be a
good approximation of f .
Algorithms that are not sensitive to scaling are preferable, because they can handle
poor problem formulations in a more robust fashion. In designing complete algorithms, we
try to incorporate scale invariance into all aspects of the algorithm, including the line search
or trust-region strategies and convergence tests. Generally speaking, it is easier to preserve
scale invariance for line search algorithms than for trust-region algorithms.
✐
EXERCISES
✐
2.1 Compute the gradient ∇ f (x) and Hessian ∇ 2 f (x) of the Rosenbrock function
f (x) 100(x2 − x12 )2 + (1 − x1 )2 .
(2.22)
27
28
CHAPTER 2.
FUNDAMENTALS
OF
UNCONSTRAINED OPTIMIZATION
Show that x ∗ (1, 1)T is the only local minimizer of this function, and that the Hessian
matrix at that point is positive definite.
✐
2.2 Show that the function f (x) 8x1 + 12x2 + x12 − 2x22 has only one stationary
point, and that it is neither a maximum or minimum, but a saddle point. Sketch the contour
lines of f .
✐
2.3 Let a be a given n-vector, and A be a given n × n symmetric matrix. Compute the
gradient and Hessian of f 1 (x) a T x and f 2 (x) x T Ax.
✐
2.4 Write the second-order Taylor expansion (2.6) for the function cos(1/x) around
a nonzero point x, and the third-order Taylor expansion of cos(x) around any point x.
Evaluate the second expansion for the specific case of x 1.
✐
2.5 Consider the function f : IR2 → IR defined by f (x) x2 . Show that the
sequence of iterates {xk } defined by
cos k
1
xk 1 + k
2
sin k
satisfies f (xk+1 ) < f (xk ) for k 0, 1, 2, . . . . Show that every point on the unit circle
{x | x2 1} is a limit point for {xk }. Hint: Every value θ ∈ [0, 2π ] is a limit point of the
subsequence {ξk } defined by
k
ξk k(mod 2π ) k − 2π
,
2π
where the operator · denotes rounding down to the next integer.
✐
2.6 Prove that all isolated local minimizers are strict. (Hint: Take an isolated local
minimizer x ∗ and a neighborhood N . Show that for any x ∈ N , x x ∗ we must have
f (x) > f (x ∗ ).)
✐
2.7 Suppose that f (x) x T Qx, where Q is an n × n symmetric positive semidefinite
matrix. Show using the definition (1.4) that f (x) is convex on the domain IRn . Hint: It may
be convenient to prove the following equivalent inequality:
f (y + α(x − y)) − α f (x) − (1 − α) f (y) ≤ 0,
for all α ∈ [0, 1] and all x, y ∈ IRn .
✐
2.8 Suppose that f is a convex function. Show that the set of global minimizers of f
is a convex set.
2.2.
OVERVIEW
OF
ALGORITHMS
2
2.9 Consider the function f (x1 , x2 ) x1 + x22 . At the point x T (1, 0) we
consider the search direction p T (−1, 1). Show that p is a descent direction and find all
minimizers of the problem (2.10).
✐
2.10 Suppose that f˜(z) f (x), where x Sz + s for some S ∈ IRn×n and s ∈ IRn .
Show that
✐
∇ f˜(z) S T ∇ f (x),
∇ 2 f˜(z) S T ∇ 2 f (x)S.
(Hint: Use the chain rule to express d f˜/dz j in terms of d f /d xi and d xi /dz j for all
i, j 1, 2, . . . , n.)
✐
2.11 Show that the symmetric rank-one update (2.18) and the BFGS update (2.19)
are scale-invariant if the initial Hessian approximations B0 are chosen appropriately. That
is, using the notation of the previous exercise, show that if these methods are applied to
f (x) starting from x0 Sz 0 + s with initial Hessian B0 , and to f˜(z) starting from z 0 with
initial Hessian S T B0 S, then all iterates are related by xk Sz k + s. (Assume for simplicity
that the methods take unit step lengths.)
✐
2.12 Suppose that a function f of two variables is poorly scaled at the solution x ∗ .
Write two Taylor expansions of f around x ∗ —one along each coordinate direction—and
use them to show that the Hessian ∇ 2 f (x ∗ ) is ill-conditioned.
✐
2.13 (For this and the following three questions, refer to the material on “Rates of
Convergence” in Section A.2 of the Appendix.) Show that the sequence xk 1/k is not
Q-linearly convergent, though it does converge to zero. (This is called sublinear convergence.)
✐
2.14 Show that the sequence xk 1 + (0.5)2 is Q-quadratically convergent to 1.
✐
2.15 Does the sequence xk 1/k! converge Q-superlinearly? Q-quadratically?
✐
2.16 Consider the sequence {xk } defined by
k
xk 2k
1
,
k even,
(xk−1 )/k,
k odd.
4
Is this sequence Q-superlinearly convergent? Q-quadratically convergent? R-quadratically
convergent?
29
This is pag
Printer: O
CHAPTER
3
Line Search
Methods
Each iteration of a line search method computes a search direction pk and then decides how
far to move along that direction. The iteration is given by
xk+1 xk + αk pk ,
(3.1)
where the positive scalar αk is called the step length. The success of a line search method
depends on effective choices of both the direction pk and the step length αk .
Most line search algorithms require pk to be a descent direction—one for which
pkT ∇ f k < 0—because this property guarantees that the function f can be reduced along
3.1.
STEP LENGTH
this direction, as discussed in the previous chapter. Moreover, the search direction often has
the form
pk −Bk−1 ∇ f k ,
(3.2)
where Bk is a symmetric and nonsingular matrix. In the steepest descent method, Bk is
simply the identity matrix I , while in Newton’s method, Bk is the exact Hessian ∇ 2 f (xk ).
In quasi-Newton methods, Bk is an approximation to the Hessian that is updated at every
iteration by means of a low-rank formula. When pk is defined by (3.2) and Bk is positive
definite, we have
pkT ∇ f k −∇ f kT Bk−1 ∇ f k < 0,
and therefore pk is a descent direction.
In this chapter, we discuss how to choose αk and pk to promote convergence from
remote starting points. We also study the rate of convergence of steepest descent, quasiNewton, and Newton methods. Since the pure Newton iteration is not guaranteed to produce
descent directions when the current iterate is not close to a solution, we discuss modifications
in Section 3.4 that allow it to start from any initial point.
We now give careful consideration to the choice of the step-length parameter αk .
3.1
STEP LENGTH
In computing the step length αk , we face a tradeoff. We would like to choose αk to give
a substantial reduction of f , but at the same time we do not want to spend too much
time making the choice. The ideal choice would be the global minimizer of the univariate
function φ(·) defined by
φ(α) f (xk + αpk ),
α > 0,
(3.3)
but in general, it is too expensive to identify this value (see Figure 3.1). To find even a local
minimizer of φ to moderate precision generally requires too many evaluations of the objective function f and possibly the gradient ∇ f . More practical strategies perform an inexact
line search to identify a step length that achieves adequate reductions in f at minimal cost.
Typical line search algorithms try out a sequence of candidate values for α, stopping to
accept one of these values when certain conditions are satisfied. The line search is done in two
stages: A bracketing phase finds an interval containing desirable step lengths, and a bisection
or interpolation phase computes a good step length within this interval. Sophisticated line
search algorithms can be quite complicated, so we defer a full description until Section 3.5.
31
32
CHAPTER 3.
LINE SEARCH METHODS
φ (α)
first local
minimizer
α
first
stationary
point
Figure 3.1
global minimizer
The ideal step length is the global minimizer.
We now discuss various termination conditions for line search algorithms and show
that effective step lengths need not lie near minimizers of the univariate function φ(α)
defined in (3.3).
A simple condition we could impose on αk is to require a reduction in f , that is,
f (xk + αk pk ) < f (xk ). That this requirement is not enough to produce convergence to
x ∗ is illustrated in Figure 3.2, for which the minimum function value is f ∗ −1, but a
sequence of iterates {xk } for which f (xk ) 5/k, k 0, 1, . . . yields a decrease at each
iteration but has a limiting function value of zero. The insufficient reduction in f at each
step causes it to fail to converge to the minimizer of this convex function. To avoid this
behavior we need to enforce a sufficient decrease condition, a concept we discuss next.
f( x )
x
x1
Figure 3.2
Insufficient reduction in f .
x
2
x
0
3.1.
STEP LENGTH
THE WOLFE CONDITIONS
A popular inexact line search condition stipulates that αk should first of all give
sufficient decrease in the objective function f , as measured by the following inequality:
f (xk + αpk ) ≤ f (xk ) + c1 α∇ f kT pk ,
(3.4)
for some constant c1 ∈ (0, 1). In other words, the reduction in f should be proportional to
both the step length αk and the directional derivative ∇ f kT pk . Inequality (3.4) is sometimes
called the Armijo condition.
The sufficient decrease condition is illustrated in Figure 3.3. The right-hand-side of
(3.4), which is a linear function, can be denoted by l(α). The function l(·) has negative slope
c1 ∇ f kT pk , but because c1 ∈ (0, 1), it lies above the graph of φ for small positive values of
α. The sufficient decrease condition states that α is acceptable only if φ(α) ≤ l(α). The
intervals on which this condition is satisfied are shown in Figure 3.3. In practice, c1 is chosen
to be quite small, say c1 10−4 .
The sufficient decrease condition is not enough by itself to ensure that the algorithm
makes reasonable progress because, as we see from Figure 3.3, it is satisfied for all sufficiently
small values of α. To rule out unacceptably short steps we introduce a second requirement,
called the curvature condition, which requires αk to satisfy
∇ f (xk + αk pk )T pk ≥ c2 ∇ f kT pk ,
(3.5)
for some constant c2 ∈ (c1 , 1), where c1 is the constant from (3.4). Note that the left-handside is simply the derivative φ (αk ), so the curvature condition ensures that the slope of φ at
αk is greater than c2 times the initial slope φ (0). This makes sense because if the slope φ (α)
φ (α) = f(xk+ α pk )
l( α)
α
acceptable
Figure 3.3
Sufficient decrease condition.
acceptable
33
34
CHAPTER 3.
LINE SEARCH METHODS
φ (α) =f(x k+α pk )
desired
slope
tangent
α
acceptable
Figure 3.4
acceptable
The curvature condition.
is strongly negative, we have an indication that we can reduce f significantly by moving
further along the chosen direction.
On the other hand, if φ (αk ) is only slightly negative or even positive, it is a sign that
we cannot expect much more decrease in f in this direction, so it makes sense to terminate
the line search. The curvature condition is illustrated in Figure 3.4. Typical values of c2 are
0.9 when the search direction pk is chosen by a Newton or quasi-Newton method, and 0.1
when pk is obtained from a nonlinear conjugate gradient method.
The sufficient decrease and curvature conditions are known collectively as the Wolfe
conditions. We illustrate them in Figure 3.5 and restate them here for future reference:
f (xk + αk pk ) ≤ f (xk ) + c1 αk ∇ f kT pk ,
∇ f (xk + αk pk ) pk ≥ c2 ∇
T
f kT
pk ,
(3.6a)
(3.6b)
with 0 < c1 < c2 < 1.
A step length may satisfy the Wolfe conditions without being particularly close to a
minimizer of φ, as we show in Figure 3.5. We can, however, modify the curvature condition
to force αk to lie in at least a broad neighborhood of a local minimizer or stationary point
of φ. The strong Wolfe conditions require αk to satisfy
f (xk + αk pk ) ≤ f (xk ) + c1 αk ∇ f kT pk ,
|∇ f (xk + αk pk ) pk | ≤ c2 |∇
T
f kT
pk |,
(3.7a)
(3.7b)
with 0 < c1 < c2 < 1. The only difference with the Wolfe conditions is that we no longer
allow the derivative φ (αk ) to be too positive. Hence, we exclude points that are far from
stationary points of φ.
3.1.
STEP LENGTH
φ (α ) = f(x k + α pk )
line of sufficient
decrease
l(α )
desired
slope
α
acceptable
Figure 3.5
acceptable
Step lengths satisfying the Wolfe conditions.
It is not difficult to prove that there exist step lengths that satisfy the Wolfe conditions
for every function f that is smooth and bounded below.
Lemma 3.1.
Suppose that f : IRn → IR is continuously differentiable. Let pk be a descent direction at
xk , and assume that f is bounded below along the ray {xk + αpk |α > 0}. Then if 0 < c1 <
c2 < 1, there exist intervals of step lengths satisfying the Wolfe conditions (3.6) and the strong
Wolfe conditions (3.7).
PROOF. Note that φ(α) f (xk + αpk ) is bounded below for all α > 0. Since 0 < c1 < 1,
the line l(α) f (xk ) + αc1 ∇ f kT pk is unbounded below and must therefore intersect the
graph of φ at least once. Let α > 0 be the smallest intersecting value of α, that is,
f (xk + α pk ) f (xk ) + α c1 ∇ f kT pk .
(3.8)
The sufficient decrease condition (3.6a) clearly holds for all step lengths less than α .
By the mean value theorem (see (A.55)), there exists α ∈ (0, α ) such that
f (xk + α pk ) − f (xk ) α ∇ f (xk + α pk )T pk .
(3.9)
By combining (3.8) and (3.9), we obtain
∇ f (xk + α pk )T pk c1 ∇ f kT pk > c2 ∇ f kT pk ,
(3.10)
since c1 < c2 and ∇ f kT pk < 0. Therefore, α satisfies the Wolfe conditions (3.6), and the
inequalities hold strictly in both (3.6a) and (3.6b). Hence, by our smoothness assumption
on f , there is an interval around α for which the Wolfe conditions hold. Moreover, since
35
36
CHAPTER 3.
LINE SEARCH METHODS
the term in the left-hand side of (3.10) is negative, the strong Wolfe conditions (3.7) hold in
the same interval.
The Wolfe conditions are scale-invariant in a broad sense: Multiplying the objective
function by a constant or making an affine change of variables does not alter them. They can
be used in most line search methods, and are particularly important in the implementation
of quasi-Newton methods, as we see in Chapter 6.
THE GOLDSTEIN CONDITIONS
Like the Wolfe conditions, the Goldstein conditions ensure that the step length α
achieves sufficient decrease but is not too short. The Goldstein conditions can also be stated
as a pair of inequalities, in the following way:
f (xk ) + (1 − c)αk ∇ f kT pk ≤ f (xk + αk pk ) ≤ f (xk ) + cαk ∇ f kT pk ,
(3.11)
with 0 < c < 1/2. The second inequality is the sufficient decrease condition (3.4), whereas
the first inequality is introduced to control the step length from below; see Figure 3.6
A disadvantage of the Goldstein conditions vis-à-vis the Wolfe conditions is that the
first inequality in (3.11) may exclude all minimizers of φ. However, the Goldstein and Wolfe
conditions have much in common, and their convergence theories are quite similar. The
Goldstein conditions are often used in Newton-type methods but are not well suited for
quasi-Newton methods that maintain a positive definite Hessian approximation.
φ ( α ) = f(x k+ α pk )
α c f kTp
k
α (1 _ c) fkTpk
α
acceptable steplengths
Figure 3.6
The Goldstein conditions.
3.2.
CONVERGENCE
OF
LINE SEARCH METHODS
SUFFICIENT DECREASE AND BACKTRACKING
We have mentioned that the sufficient decrease condition (3.6a) alone is not sufficient
to ensure that the algorithm makes reasonable progress along the given search direction.
However, if the line search algorithm chooses its candidate step lengths appropriately, by
using a so-called backtracking approach, we can dispense with the extra condition (3.6b)
and use just the sufficient decrease condition to terminate the line search procedure. In its
most basic form, backtracking proceeds as follows.
Algorithm 3.1 (Backtracking Line Search).
Choose ᾱ > 0, ρ ∈ (0, 1), c ∈ (0, 1); Set α ← ᾱ;
repeat until f (xk + αpk ) ≤ f (xk ) + cα∇ f kT pk
α ← ρα;
end (repeat)
Terminate with αk α.
In this procedure, the initial step length ᾱ is chosen to be 1 in Newton and quasiNewton methods, but can have different values in other algorithms such as steepest descent
or conjugate gradient. An acceptable step length will be found after a finite number of
trials, because αk will eventually become small enough that the sufficient decrease condition
holds (see Figure 3.3). In practice, the contraction factor ρ is often allowed to vary at each
iteration of the line search. For example, it can be chosen by safeguarded interpolation, as
we describe later. We need ensure only that at each iteration we have ρ ∈ [ρlo , ρhi ], for some
fixed constants 0 < ρlo < ρhi < 1.
The backtracking approach ensures either that the selected step length αk is some fixed
value (the initial choice ᾱ), or else that it is short enough to satisfy the sufficient decrease
condition but not too short. The latter claim holds because the accepted value αk is within
a factor ρ of the previous trial value, αk /ρ, which was rejected for violating the sufficient
decrease condition, that is, for being too long.
This simple and popular strategy for terminating a line search is well suited for Newton
methods but is less appropriate for quasi-Newton and conjugate gradient methods.
3.2
CONVERGENCE OF LINE SEARCH METHODS
To obtain global convergence, we must not only have well chosen step lengths but also well
chosen search directions pk . We discuss requirements on the search direction in this section,
focusing on one key property: the angle θk between pk and the steepest descent direction
−∇ f k , defined by
cos θk −∇ f kT pk
.
∇ f k pk (3.12)
37
38
CHAPTER 3.
LINE SEARCH METHODS
The following theorem, due to Zoutendijk, has far-reaching consequences. It quantifies
the effect of properly chosen step lengths αk , and shows, for example, that the steepest descent
method is globally convergent. For other algorithms, it describes how far pk can deviate
from the steepest descent direction and still produce a globally convergent iteration. Various
line search termination conditions can be used to establish this result, but for concreteness
we will consider only the Wolfe conditions (3.6). Though Zoutendijk’s result appears at first
to be technical and obscure, its power will soon become evident.
Theorem 3.2.
Consider any iteration of the form (3.1), where pk is a descent direction and αk satisfies
the Wolfe conditions (3.6). Suppose that f is bounded below in IRn and that f is continuously
def
differentiable in an open set N containing the level set L {x : f (x) ≤ f (x0 )}, where x0 is
the starting point of the iteration. Assume also that the gradient ∇ f is Lipschitz continuous on
N , that is, there exists a constant L > 0 such that
∇ f (x) − ∇ f (x̃) ≤ Lx − x̃,
for all x, x̃ ∈ N .
(3.13)
Then
cos2 θk ∇ f k 2 < ∞.
k≥0
PROOF. From (3.6b) and (3.1) we have that
(∇ f k+1 − ∇ f k )T pk ≥ (c2 − 1)∇ f kT pk ,
while the Lipschitz condition (3.13) implies that
(∇ f k+1 − ∇ f k )T pk ≤ αk L pk 2 .
By combining these two relations, we obtain
αk ≥
c2 − 1 ∇ f kT pk
.
L
pk 2
By substituting this inequality into the first Wolfe condition (3.6a), we obtain
f k+1 ≤ f k − c1
1 − c2 (∇ f kT pk )2
.
L
pk 2
From the definition (3.12), we can write this relation as
f k+1 ≤ f k − c cos2 θk ∇ f k 2 ,
(3.14)
3.2.
CONVERGENCE
OF
LINE SEARCH METHODS
where c c1 (1 − c2 )/L. By summing this expression over all indices less than or equal to
k, we obtain
f k+1 ≤ f 0 − c
k
cos2 θ j ∇ f j 2 .
(3.15)
j0
Since f is bounded below, we have that f 0 − f k+1 is less than some positive constant, for all
k. Hence, by taking limits in (3.15), we obtain
∞
cos2 θk ∇ f k 2 < ∞,
k0
which concludes the proof.
Similar results to this theorem hold when the Goldstein conditions (3.11) or strong
Wolfe conditions (3.7) are used in place of the Wolfe conditions. For all these strategies, the
step length selection implies inequality (3.14), which we call the Zoutendijk condition.
Note that the assumptions of Theorem 3.2 are not too restrictive. If the function f were
not bounded below, the optimization problem would not be well defined. The smoothness
assumption—Lipschitz continuity of the gradient—is implied by many of the smoothness
conditions that are used in local convergence theorems (see Chapters 6 and 7) and are often
satisfied in practice.
The Zoutendijk condition (3.14) implies that
cos2 θk ∇ f k 2 → 0.
(3.16)
This limit can be used in turn to derive global convergence results for line search algorithms.
If our method for choosing the search direction pk in the iteration (3.1) ensures that
the angle θk defined by (3.12) is bounded away from 90◦ , there is a positive constant δ such
that
cos θk ≥ δ > 0,
for all k.
(3.17)
It follows immediately from (3.16) that
lim ∇ f k 0.
k→∞
(3.18)
In other words, we can be sure that the gradient norms ∇ f k converge to zero, provided
that the search directions are never too close to orthogonality with the gradient. In particular,
the method of steepest descent (for which the search direction pk is parallel to the negative
39
40
CHAPTER 3.
LINE SEARCH METHODS
gradient) produces a gradient sequence that converges to zero, provided that it uses a line
search satisfying the Wolfe or Goldstein conditions.
We use the term globally convergent to refer to algorithms for which the property
(3.18) is satisfied, but note that this term is sometimes used in other contexts to mean
different things. For line search methods of the general form (3.1), the limit (3.18) is the
strongest global convergence result that can be obtained: We cannot guarantee that the
method converges to a minimizer, but only that it is attracted by stationary points. Only
by making additional requirements on the search direction pk —by introducing negative
curvature information from the Hessian ∇ 2 f (xk ), for example—can we strengthen these
results to include convergence to a local minimum. See the Notes and References at the end
of this chapter for further discussion of this point.
Consider now the Newton-like method (3.1), (3.2) and assume that the matrices Bk
are positive definite with a uniformly bounded condition number. That is, there is a constant
M such that
Bk Bk−1 ≤ M,
for all k.
It is easy to show from the definition (3.12) that
cos θk ≥ 1/M
(3.19)
(see Exercise 3.5). By combining this bound with (3.16) we find that
lim ∇ f k 0.
k→∞
(3.20)
Therefore, we have shown that Newton and quasi-Newton methods are globally convergent
if the matrices Bk have a bounded condition number and are positive definite (which is
needed to ensure that pk is a descent direction), and if the step lengths satisfy the Wolfe
conditions.
For some algorithms, such as conjugate gradient methods, we will be able to prove
the limit (3.18), but only the weaker result
lim inf ∇ f k 0.
k→∞
(3.21)
In other words, just a subsequence of the gradient norms ∇ f k j converges to zero, rather
than the whole sequence (see Appendix A). This result, too, can be proved by using Zoutendijk’s condition (3.14), but instead of a constructive proof, we outline a proof by
contradiction. Suppose that (3.21) does not hold, so that the gradients remain bounded
away from zero, that is, there exists γ > 0 such that
∇ f k ≥ γ ,
for all k sufficiently large.
(3.22)
3.3.
RATE
OF
CONVERGENCE
Then from (3.16) we conclude that
cos θk → 0,
(3.23)
that is, the entire sequence {cos θk } converges to 0. To establish (3.21), therefore, it is enough
to show that a subsequence {cos θk j } is bounded away from zero. We will use this strategy in
Chapter 5 to study the convergence of nonlinear conjugate gradient methods.
By applying this proof technique, we can prove global convergence in the sense of
(3.20) or (3.21) for a general class of algorithms. Consider any algorithm for which (i) every
iteration produces a decrease in the objective function, and (ii) every mth iteration is a
steepest descent step, with step length chosen to satisfy the Wolfe or Goldstein conditions.
Then, since cos θk 1 for the steepest descent steps, the result (3.21) holds. Of course, we
would design the algorithm so that it does something “better" than steepest descent at the
other m − 1 iterates. The occasional steepest descent steps may not make much progress,
but they at least guarantee overall global convergence.
Note that throughout this section we have used only the fact that Zoutendijk’s condition implies the limit (3.16). In later chapters we will make use of the bounded sum condition
(3.14), which forces the sequence {cos2 θk ∇ f k 2 } to converge to zero at a sufficiently rapid
rate.
3.3
RATE OF CONVERGENCE
It would seem that designing optimization algorithms with good convergence properties is
easy, since all we need to ensure is that the search direction pk does not tend to become
orthogonal to the gradient ∇ f k , or that steepest descent steps are taken regularly. We could
simply compute cos θk at every iteration and turn pk toward the steepest descent direction if
cos θk is smaller than some preselected constant δ > 0. Angle tests of this type ensure global
convergence, but they are undesirable for two reasons. First, they may impede a fast rate of
convergence, because for problems with an ill-conditioned Hessian, it may be necessary to
produce search directions that are almost orthogonal to the gradient, and an inappropriate
choice of the parameter δ may cause such steps to be rejected. Second, angle tests destroy
the invariance properties of quasi-Newton methods.
Algorithmic strategies that achieve rapid convergence can sometimes conflict with
the requirements of global convergence, and vice versa. For example, the steepest descent
method is the quintessential globally convergent algorithm, but it is quite slow in practice,
as we shall see below. On the other hand, the pure Newton iteration converges rapidly when
started close enough to a solution, but its steps may not even be descent directions away
from the solution. The challenge is to design algorithms that incorporate both properties:
good global convergence guarantees and a rapid rate of convergence.
We begin our study of convergence rates of line search methods by considering the
most basic approach of all: the steepest descent method.
41
42
CHAPTER 3.
Figure 3.7
LINE SEARCH METHODS
Steepest descent steps.
CONVERGENCE RATE OF STEEPEST DESCENT
We can learn much about the steepest descent method by considering the ideal case, in
which the objective function is quadratic and the line searches are exact. Let us suppose that
f (x) 12 x T Qx − b T x,
(3.24)
where Q is symmetric and positive definite. The gradient is given by ∇ f (x) Qx − b and
the minimizer x ∗ is the unique solution of the linear system Qx b.
It is easy to compute the step length αk that minimizes f (xk −α∇ f k ). By differentiating
the function
1
f (xk − α∇ f k ) (xk − α∇ f k )T Q(xk − α∇ f k ) − b T (xk − α∇ f k )
2
with respect to α, and setting the derivative to zero, we obtain
αk ∇ f kT ∇ f k
.
∇ f kT Q∇ f k
(3.25)
If we use this exact minimizer αk , the steepest descent iteration for (3.24) is given by
xk+1 xk −
∇ f kT ∇ f k
∇ f kT Q∇ f k
∇ fk .
(3.26)
Since ∇ f k Qxk − b, this equation yields a closed-form expression for xk+1 in terms of xk .
In Figure 3.7 we plot a typical sequence of iterates generated by the steepest descent method
on a two-dimensional quadratic objective function. The contours of f are ellipsoids whose
3.3.
RATE
OF
CONVERGENCE
axes lie along the orthogonal eigenvectors of Q. Note that the iterates zigzag toward the
solution.
To quantify the rate of convergence we introduce the weighted norm x2Q x T Qx.
By using the relation Qx ∗ b, we can show that
1
x
2
− x ∗ 2Q f (x) − f (x ∗ ),
(3.27)
so this norm measures the difference between the current objective value and the optimal
value. By using the equality (3.26) and noting that ∇ f k Q(xk − x ∗ ), we can derive the
equality
2
T
∇
f
∇
f
k
k
xk − x ∗ 2Q
xk+1 − x ∗ 2Q 1 − T
∇ f k Q∇ f k ∇ f kT Q −1 ∇ f k
(3.28)
(see Exercise 3.7). This expression describes the exact decrease in f at each iteration, but
since the term inside the brackets is difficult to interpret, it is more useful to bound it in
terms of the condition number of the problem.
Theorem 3.3.
When the steepest descent method with exact line searches (3.26) is applied to the strongly
convex quadratic function (3.24), the error norm (3.27) satisfies
xk+1 −
x ∗ 2Q
≤
λn − λ1
λn + λ1
2
xk − x ∗ 2Q ,
(3.29)
where 0 < λ1 ≤ λ2 ≤ · · · ≤ λn are the eigenvalues of Q.
The proof of this result is given by Luenberger [195]. The inequalities (3.29) and (3.27)
show that the function values f k converge to the minimum f ∗ at a linear rate. As a special
case of this result, we see that convergence is achieved in one iteration if all the eigenvalues
are equal. In this case, Q is a multiple of the identity matrix, so the contours in Figure 3.7
are circles and the steepest descent direction always points at the solution. In general, as the
condition number κ(Q) λn /λ1 increases, the contours of the quadratic become more
elongated, the zigzagging in Figure 3.7 becomes more pronounced, and (3.29) implies that
the convergence degrades. Even though (3.29) is a worst-case bound, it gives an accurate
indication of the behavior of the algorithm when n > 2.
The rate-of-convergence behavior of the steepest descent method is essentially the
same on general nonlinear objective functions. In the following result we assume that the
step length is the global minimizer along the search direction.
Theorem 3.4.
Suppose that f : IRn → IR is twice continuously differentiable, and that the iterates
generated by the steepest-descent method with exact line searches converge to a point x ∗ at
43
44
CHAPTER 3.
LINE SEARCH METHODS
which the Hessian matrix ∇ 2 f (x ∗ ) is positive definite. Let r be any scalar satisfying
λn − λ1
r∈
,1 ,
λn + λ1
where λ1 ≤ λ2 ≤ · · · ≤ λn are the eigenvalues of ∇ 2 f (x ∗ ). Then for all k sufficiently large, we
have
f (xk+1 ) − f (x ∗ ) ≤ r 2 [ f (xk ) − f (x ∗ )].
In general, we cannot expect the rate of convergence to improve if an inexact line
search is used. Therefore, Theorem 3.4 shows that the steepest descent method can have an
unacceptably slow rate of convergence, even when the Hessian is reasonably well conditioned.
For example, if κ(Q) 800, f (x1 ) 1, and f (x ∗ ) 0, Theorem 3.4 suggests that the
function value will still be about 0.08 after one thousand iterations of the steepest descent
method with exact line search.
NEWTON’S METHOD
We now consider the Newton iteration, for which the search is given by
pkN −∇ 2 f k−1 ∇ f k .
(3.30)
Since the Hessian matrix ∇ 2 f k may not always be positive definite, pkN may not always
be a descent direction, and many of the ideas discussed so far in this chapter no longer
apply. In Section 3.4 and Chapter 4 we will describe two approaches for obtaining a globally
convergent iteration based on the Newton step: a line search approach, in which the Hessian
∇ 2 f k is modified, if necessary, to make it positive definite and thereby yield descent, and a
trust region approach, in which ∇ 2 f k is used to form a quadratic model that is minimized
in a ball around the current iterate xk .
Here we discuss just the local rate-of-convergence properties of Newton’s method.
We know that for all x in the vicinity of a solution point x ∗ such that ∇ 2 f (x ∗ ) is positive
definite, the Hessian ∇ 2 f (x) will also be positive definite. Newton’s method will be well
defined in this region and will converge quadratically, provided that the step lengths αk are
eventually always 1.
Theorem 3.5.
Suppose that f is twice differentiable and that the Hessian ∇ 2 f (x) is Lipschitz continuous
(see (A.42)) in a neighborhood of a solution x ∗ at which the sufficient conditions (Theorem 2.4)
are satisfied. Consider the iteration xk+1 xk + pk , where pk is given by (3.30). Then
(i) if the starting point x0 is sufficiently close to x ∗ , the sequence of iterates converges to x ∗ ;
(ii) the rate of convergence of {xk } is quadratic; and
(iii) the sequence of gradient norms {∇ f k } converges quadratically to zero.
3.3.
RATE
OF
CONVERGENCE
PROOF. From the definition of the Newton step and the optimality condition ∇ f ∗ 0 we
have that
xk + pkN − x ∗ xk − x ∗ − ∇ 2 f k−1 ∇ f k
∇ 2 f k−1 ∇ 2 f k (xk − x ∗ ) − (∇ f k − ∇ f ∗ ) .
(3.31)
Since Taylor’s theorem (Theorem 2.1) tells us that
1
∇ 2 f (xk + t(x ∗ − xk ))(xk − x ∗ ) dt,
∇ fk − ∇ f∗ 0
we have
2
∇ f (xk )(xk − x ∗ ) − (∇ f k − ∇ f (x ∗ ))
1
2
2
∗
∗
∇ f (xk ) − ∇ f (xk + t(x − xk )) (xk − x ) dt 0
1
∇ 2 f (xk ) − ∇ 2 f (xk + t(x ∗ − xk )) xk − x ∗ dt
0
1
∗ 2
Lt dt 12 Lxk − x ∗ 2 ,
≤ xk − x ≤
(3.32)
0
where L is the Lipschitz constant for ∇ 2 f (x) for x near x ∗ . Since ∇ 2 f (x ∗ ) is nonsingular,
there is a radius r > 0 such that ∇ 2 f k−1 ≤ 2∇ 2 f (x ∗ )−1 for all xk with xk − x ∗ ≤ r .
By substituting in (3.31) and (3.32), we obtain
xk + pkN − x ∗ ≤ L∇ 2 f (x ∗ )−1 xk − x ∗ 2 L̃xk − x ∗ 2 ,
(3.33)
where L̃ L∇ 2 f (x ∗ )−1 . Choosing x0 so that x0 − x ∗ ≤ min(r, 1/(2 L̃)), we can use
this inequality inductively to deduce that the sequence converges to x ∗ , and the rate of
convergence is quadratic.
By using the relations xk+1 − xk pkN and ∇ f k + ∇ 2 f k pkN 0, we obtain that
∇ f (xk+1 ) ∇ f (xk+1 ) − ∇ f k − ∇ 2 f (xk ) pkN 1
2
2
N
N
∇ f (xk + t pk )(xk+1 − xk ) dt − ∇ f (xk ) pk 0
1
≤
0
∇ 2 f (xk + t p N ) − ∇ 2 f (xk ) p N dt
k
k
≤ 12 L pkN 2
≤ 12 L∇ 2 f (xk )−1 2 ∇ f k 2
≤ 2L∇ 2 f (x ∗ )−1 2 ∇ f k 2 ,
proving that the gradient norms converge to zero quadratically.
45
46
CHAPTER 3.
LINE SEARCH METHODS
As the iterates generated by Newton’s method approach the solution, the Wolfe (or
Goldstein) conditions will accept the step length αk 1 for all large k. This observation
follows from Theorem 3.6 below. Indeed, when the search direction is given by Newton’s
method, the limit (3.35) is satisfied—the ratio is zero for all k! Implementations of Newton’s
method using these line search conditions, and in which the line search always tries the unit
step length first, will set αk 1 for all large k and attain a local quadratic rate of convergence.
QUASI-NEWTON METHODS
Suppose now that the search direction has the form
pk −Bk−1 ∇ f k ,
(3.34)
where the symmetric and positive definite matrix Bk is updated at every iteration by a
quasi-Newton updating formula. We already encountered one quasi-Newton formula, the
BFGS formula, in Chapter 2; others will be discussed in Chapter 6. We assume here that the
step length αk is computed by an inexact line search that satisfies the Wolfe or strong Wolfe
conditions, with the same proviso mentioned above for Newton’s method: The line search
algorithm will always try the step length α 1 first, and will accept this value if it satisfies
the Wolfe conditions. (We could enforce this condition by setting ᾱ 1 in Algorithm 3.1,
for example.) This implementation detail turns out to be crucial in obtaining a fast rate of
convergence.
The following result shows that if the search direction of a quasi-Newton method
approximates the Newton direction well enough, then the unit step length will satisfy the
Wolfe conditions as the iterates converge to the solution. It also specifies a condition that
the search direction must satisfy in order to give rise to a superlinearly convergent iteration.
To bring out the full generality of this result, we state it first in terms of a general descent
iteration, and then examine its consequences for quasi-Newton and Newton methods.
Theorem 3.6.
Suppose that f : IRn → IR is twice continuously differentiable. Consider the iteration
xk+1 xk + αk pk , where pk is a descent direction and αk satisfies the Wolfe conditions (3.6)
with c1 ≤ 1/2. If the sequence {xk } converges to a point x ∗ such that ∇ f (x ∗ ) 0 and ∇ 2 f (x ∗ )
is positive definite, and if the search direction satisfies
∇ f k + ∇ 2 f k pk 0,
k→∞
pk lim
then
(i) the step length αk 1 is admissible for all k greater than a certain index k0 ; and
(ii) if αk 1 for all k > k0 , {xk } converges to x ∗ superlinearly.
(3.35)
3.3.
RATE
OF
CONVERGENCE
It is easy to see that if c1 > 1/2, then the line search would exclude the minimizer of
a quadratic, and unit step lengths may not be admissible.
If pk is a quasi-Newton search direction of the form (3.34), then (3.35) is equivalent to
(Bk − ∇ 2 f (x ∗ )) pk 0.
k→∞
pk lim
(3.36)
Hence, we have the surprising (and delightful) result that a superlinear convergence rate
can be attained even if the sequence of quasi-Newton matrices Bk does not converge to
∇ 2 f (x ∗ ); it suffices that the Bk become increasingly accurate approximations to ∇ 2 f (x ∗ )
along the search directions pk . Importantly, condition (3.36) is both necessary and sufficient
for the superlinear convergence of quasi-Newton methods.
Theorem 3.7.
Suppose that f : IRn → IR is twice continuously differentiable. Consider the iteration
xk+1 xk + pk (that is, the step length αk is uniformly 1) and that pk is given by (3.34). Let us
assume also that {xk } converges to a point x ∗ such that ∇ f (x ∗ ) 0 and ∇ 2 f (x ∗ ) is positive
definite. Then {xk } converges superlinearly if and only if (3.36) holds.
PROOF. We first show that (3.36) is equivalent to
pk − pkN o( pk ),
(3.37)
where pkN −∇ 2 f k−1 ∇ f k is the Newton step. Assuming that (3.36) holds, we have that
pk − pkN ∇ 2 f k−1 (∇ 2 f k pk + ∇ f k )
∇ 2 f k−1 (∇ 2 f k − Bk ) pk
O((∇ 2 f k − Bk ) pk )
o( pk ),
where we have used the fact that ∇ 2 f k−1 is bounded above for xk sufficiently close to x ∗ ,
since the limiting Hessian ∇ 2 f (x ∗ ) is positive definite. The converse follows readily if we
multiply both sides of (3.37) by ∇ 2 f k and recall (3.34).
By combining (3.33) and (3.37), we obtain that
xk + pk − x ∗ ≤ xk + pkN − x ∗ + pk − pkN O(xk − x ∗ 2 ) + o( pk ).
A simple manipulation of this inequality reveals that pk O(xk − x ∗ ), so we obtain
xk + pk − x ∗ ≤ o(xk − x ∗ ),
giving the superlinear convergence result.
47
48
CHAPTER 3.
LINE SEARCH METHODS
We will see in Chapter 6 that quasi-Newton methods normally satisfy condition (3.36)
and are therefore superlinearly convergent.
3.4
NEWTON’S METHOD WITH HESSIAN MODIFICATION
Away from the solution, the Hessian matrix ∇ 2 f (x) may not be positive definite, so the
Newton direction pkN defined by
∇ 2 f (xk ) pkN −∇ f (xk )
(3.38)
(see (3.30)) may not be a descent direction. We now describe an approach to overcome this
difficulty when a direct linear algebra technique, such as Gaussian elimination, is used to
solve the Newton equations (3.38). This approach obtains the step pk from a linear system
identical to (3.38), except that the coefficient matrix is replaced with a positive definite
approximation, formed before or during the solution process. The modified Hessian is
obtained by adding either a positive diagonal matrix or a full matrix to the true Hessian
∇ 2 f (xk ). A general description of this method follows.
Algorithm 3.2 (Line Search Newton with Modification).
Given initial point x0 ;
for k 0, 1, 2, . . .
Factorize the matrix Bk ∇ 2 f (xk ) + E k , where E k 0 if ∇ 2 f (xk )
is sufficiently positive definite; otherwise, E k is chosen to
ensure that Bk is sufficiently positive definite;
Solve Bk pk −∇ f (xk );
Set xk+1 ← xk + αk pk , where αk satisfies the Wolfe, Goldstein, or
Armijo backtracking conditions;
end
Some approaches do not compute E k explicitly, but rather introduce extra steps and
tests into standard factorization procedures, modifying these procedures “on the fly” so
that the computed factors are the factors of a positive definite matrix. Strategies based on
modifying a Cholesky factorization and on modifying a symmetric indefinite factorization
of the Hessian are described in this section.
Algorithm 3.2 is a practical Newton method that can be applied from any starting
point. We can establish fairly satisfactory global convergence results for it, provided that
the strategy for choosing E k (and hence Bk ) satisfies the bounded modified factorization
property. This property is that the matrices in the sequence {Bk } have bounded condition
number whenever the sequence of Hessians {∇ 2 f (xk )} is bounded; that is,
κ(Bk ) Bk Bk−1 ≤ C,
some C > 0 and all k 0, 1, 2, . . . .
(3.39)
3.4.
NEWTON’S METHOD
WITH
HESSIAN MODIFICATION
If this property holds, global convergence of the modified line search Newton method follows
from the results of Section 3.2.
Theorem 3.8.
Let f be twice continuously differentiable on an open set D, and assume that the starting
point x0 of Algorithm 3.2 is such that the level set L {x ∈ D : f (x) ≤ f (x0 )} is compact.
Then if the bounded modified factorization property holds, we have that
lim ∇ f (xk ) 0.
k→∞
For a proof this result see [215].
We now consider the convergence rate of Algorithm 3.2. Suppose that the sequence
of iterates xk converges to a point x ∗ where ∇ 2 f (x ∗ ) is sufficiently positive definite in the
sense that the modification strategies described in the next section return the modification
E k 0 for all sufficiently large k. By Theorem 3.6, we have that αk 1 for all sufficiently
large k, so that Algorithm 3.2 reduces to a pure Newton method, and the rate of convergence
is quadratic.
For problems in which ∇ f ∗ is close to singular, there is no guarantee that the modification E k will eventually vanish, and the convergence rate may be only linear. Besides
requiring the modified matrix Bk to be well conditioned (so that Theorem 3.8 holds), we
would like the modification to be as small as possible, so that the second-order information
in the Hessian is preserved as far as possible. Naturally, we would also like the modified
factorization to be computable at moderate cost.
To set the stage for the matrix factorization techniques that will be used in Algorithm 3.2, we will begin by assuming that the eigenvalue decomposition of ∇ 2 f (xk ) is
available. This is not realistic for large-scale problems because this decomposition is generally
too expensive to compute, but it will motivate several practical modification strategies.
EIGENVALUE MODIFICATION
Consider a problem in which, at the current iterate xk , ∇ f (xk ) (1, −3, 2)T and
∇ f (xk ) diag(10, 3, −1), which is clearly indefinite. By the spectral decomposition
theorem (see Appendix A) we can define Q I and diag(λ1 , λ2 , λ3 ), and write
2
∇ 2 f (xk ) QQ T n
λi qi qiT .
(3.40)
i1
The pure Newton step—the solution of (3.38)—is pkN (−0.1, 1, 2)T , which is not a descent direction, since ∇ f (xk )T pkN > 0. One might suggest a modified strategy in which we
replace ∇ 2 f (xk ) by a positive definite approximation Bk , in which all negative eigenvalues
in ∇ 2 f (xk ) are replaced by a small positive number δ that is somewhat larger than ma√
chine precision u; say δ u. For a machine precision of 10−16 , the resulting matrix in
49
50
CHAPTER 3.
LINE SEARCH METHODS
our example is
Bk 2
λi qi qiT + δq3 q3T diag 10, 3, 10−8 ,
(3.41)
i1
which is numerically positive definite and whose curvature along the eigenvectors q1 and
q2 has been preserved. Note, however, that the search direction based on this modified
Hessian is
pk −Bk−1 ∇ f k −
≈ − 2 × 108 q3 .
2
1 1 T
qi qi ∇ f k − q3 q3T ∇ f (xk )
λ
δ
i1 i
(3.42)
For small δ, this step is nearly parallel to q3 (with relatively small contributions from q1 and
q2 ) and quite long. Although f decreases along the direction pk , its extreme length violates
the spirit of Newton’s method, which relies on a quadratic approximation of the objective
function that is valid in a neighborhood of the current iterate xk . It is therefore not clear
that this search direction is effective.
Various other modification strategies are possible. We could flip the signs of the
negative eigenvalues in (3.40), which amounts to setting δ 1 in our example. We could
set the last term in (3.42) to zero, so that the search direction has no components along
the negative curvature directions. We could adapt the choice of δ to ensure that the length
of the step is not excessive, a strategy that has the flavor of trust-region methods. As this
discussion shows, there is a great deal of freedom in devising modification strategies, and
there is currently no agreement on which strategy is best.
Setting the issue of the choice of δ aside for the moment, let us look more closely at the
process of modifying a matrix so that it becomes positive definite. The modification (3.41)
to the example matrix (3.40) can be shown to be optimal in the following sense. If A is a
symmetric matrix with spectral decomposition A QQ T , then the correction matrix
A of minimum Frobenius norm that ensures that λmin (A + A) ≥ δ is given by
0,
λi ≥ δ,
T
A Q diag (τi )Q ,
with
τi (3.43)
δ − λi , λi < δ.
Here, λmin (A) denotes the smallest eigenvalue of A, and the Frobenius norm of a matrix is
defined as A2F i,n j1 ai2j (see (A.9)). Note that A is not diagonal in general, and that
the modified matrix is given by
A + A Q( + diag(τi ))Q T .
By using a different norm we can obtain a diagonal modification. Suppose again that
A is a symmetric matrix with spectral decomposition A QQ T . A correction matrix
3.4.
NEWTON’S METHOD
WITH
HESSIAN MODIFICATION
A with minimum Euclidean norm that satisfies λmin (A + A) ≥ δ is given by
A τ I,
with
τ max(0, δ − λmin (A)).
(3.44)
The modified matrix now has the form
A + τ I,
(3.45)
which happens to have the same form as the matrix occurring in (unscaled) trust–region
methods (see Chapter 4). All the eigenvalues of (3.45) have thus been shifted, and all are
greater than δ.
These results suggest that both diagonal and nondiagonal modifications can be considered. Even though we have not answered the question of what constitutes a good
modification, various practical diagonal and nondiagonal modifications have been proposed and implemented in software. They do not make use of the spectral decomposition of
the Hessian, since it is generally too expensive to compute. Instead, they use Gaussian elimination, choosing the modifications indirectly and hoping that somehow they will produce
good steps. Numerical experience indicates that the strategies described next often (but not
always) produce good search directions.
ADDING A MULTIPLE OF THE IDENTITY
Perhaps the simplest idea is to find a scalar τ > 0 such that ∇ 2 f (xk ) + τ I is sufficiently
positive definite. From the previous discussion we know that τ must satisfy (3.44), but a good
estimate of the smallest eigenvalue of the Hessian is normally not available. The following
algorithm describes a method that tries successively larger values of τ . (Here, aii denotes a
diagonal element of A.)
Algorithm 3.3 (Cholesky with Added Multiple of the Identity).
Choose β > 0;
if mini aii > 0
set τ0 ← 0;
else
τ0 − min(aii ) + β;
end (if)
for k 0, 1, 2, . . .
Attempt to apply the Cholesky algorithm to obtain L L T A + τk I ;
if the factorization is completed successfully
stop and return L;
else
τk+1 ← max(2τk , β);
end (if)
end (for)
51
52
CHAPTER 3.
LINE SEARCH METHODS
The choice of β is heuristic; a typical value is β 10−3 . We could choose the first
nonzero shift τ0 to be proportional to be the final value of τ used in the latest Hessian
modification; see also Algorithm B.1. The strategy implemented in Algorithm 3.3 is quite
simple and may be preferable to the modified factorization techniques described next, but
it suffers from one drawback. Every value of τk requires a new factorization of A + τk I , and
the algorithm can be quite expensive if several trial values are generated. Therefore it may
be advantageous to increase τ more rapidly, say by a factor of 10 instead of 2 in the last else
clause.
MODIFIED CHOLESKY FACTORIZATION
Another approach for modifying a Hessian matrix that is not positive definite is
to perform a Cholesky factorization of ∇ 2 f (xk ), but to increase the diagonal elements
encountered during the factorization (where necessary) to ensure that they are sufficiently
positive. This modified Cholesky approach is designed to accomplish two goals: It guarantees
that the modified Cholesky factors exist and are bounded relative to the norm of the actual
Hessian, and it does not modify the Hessian if it is sufficiently positive definite.
We begin our description of this approach by briefly reviewing the Cholesky
factorization. Every symmetric positive definite matrix A can be written as
A L DL T ,
(3.46)
where L is a lower triangular matrix with unit diagonal elements and D is a diagonal matrix
with positive elements on the diagonal. By equating the elements in (3.46), column by
column, it is easy to derive formulas for computing L and D.
❏ EXAMPLE 3.1
Consider the case n 3. The equation A L DL T is given by
⎡
a11
⎢
⎣ a21
a31
a21
a22
a32
a31
⎤
⎡
1
⎥ ⎢ l
a32 ⎦ ⎢
⎣ 21
l31
a33
0
1
l32
0
⎤⎡
d1
⎥
0 ⎥⎢
⎦⎣ 0
1
0
0
0
⎤⎡
1
d2
⎥⎢
0 ⎦⎣ 0
0
d3
0
l21
l31
⎤
1
⎥
l32 ⎦ .
0
1
(The notation indicates that A is symmetric.) By equating the elements of the first column,
we have
a11 d1 ,
a21 d1l21
⇒
l21 a21 /d1 ,
a31 d1l31
⇒
l31 a31 /d1 .
3.4.
NEWTON’S METHOD
WITH
HESSIAN MODIFICATION
Proceeding with the next two columns, we obtain
2
a22 d1l21
+ d2
⇒
2
d2 a22 − d1l21
,
a32 d1l31l21 + d2l32
⇒
l32 (a32 − d1l31l21 ) /d2 ,
a33 ⇒
2
2
d3 a33 − d1l31
− d2l32
.
2
d1l31
+
2
d2l32
+ d3
❐
This procedure is generalized in the following algorithm.
Algorithm 3.4 (Cholesky Factorization, L DL T Form).
for
j 1, 2, . . . , n
j−1
c j j ← a j j − s1 ds l 2js ;
dj ← cjj;
for i j + 1, . . . , n
j−1
ci j ← ai j − s1 ds lis l js ;
li j ← ci j /d j ;
end
end
One can show (see, for example, Golub and Van Loan [136, Section 4.2.3]) that the
diagonal elements d j j are all positive whenever A is positive definite. The scalars ci j have
been introduced only to facilitate the description of the modified factorization discussed
below. We should note that Algorithm 3.4 differs a little from the standard form of the
Cholesky factorization, which produces a lower triangular matrix M such that
A M MT .
(3.47)
In fact, we can make the identification M L D 1/2 to relate M to the factors L and D
computed in Algorithm 3.4. The technique for computing M appears as Algorithm A.2 in
Appendix A.
If A is indefinite, the factorization A L DL T may not exist. Even if it does exist,
Algorithm 3.4 is numerically unstable when applied to such matrices, in the sense that the
elements of L and D can become arbitrarily large. It follows that a strategy of computing
the L DL T factorization and then modifying the diagonal after the fact to force its elements
to be positive may break down, or may result in a matrix that is drastically different from A.
Instead, we can modify the matrix A during the course of the factorization in such
a way that all elements in D are sufficiently positive, and so that the elements of D and
L are not too large. To control the quality of the modification, we choose two positive
parameters δ and β, and require that during the computation of the jth columns of L and
D in Algorithm 3.4 (that is, for each j in the outer loop of the algorithm) the following
53
54
CHAPTER 3.
LINE SEARCH METHODS
bounds be satisfied:
d j ≥ δ,
|m i j | ≤ β,
i j + 1, j + 2, . . . , n,
(3.48)
where m i j li j d j . To satisfy these bounds we only need to change one step in Algorithm 3.4: The formula for computing the diagonal element d j in Algorithm 3.4 is replaced
by
θj
d j max |c j j |,
β
2
,δ ,
with θ j max |ci j |.
j<i≤n
(3.49)
To verify that (3.48) holds, we note from Algorithm 3.4 that ci j li j d j , and therefore
|ci j |
|ci j |β
|m i j | |li j d j | ≤
≤ β,
θj
dj
for all i > j.
We note that θ j can be computed prior to d j because the elements ci j in the second
for loop of Algorithm 3.4 do not involve d j . In fact, this is the reason for introducing the
quantities ci j into the algorithm.
These observations are the basis of the modified Cholesky algorithm described in detail
in Gill, Murray, and Wright [130], which introduces symmetric interchanges of rows and
columns to try to reduce the size of the modification. If P denotes the permutation matrix
associated with the row and column interchanges, the algorithm produces the Cholesky
factorization of the permuted, modified matrix P A P T + E, that is,
P A P T + E L DL T M M T ,
(3.50)
where E is a nonnegative diagonal matrix that is zero if A is sufficiently positive definite.
One can show (Moré and Sorensen [215]) that the matrices Bk obtained by this modified
Cholesky algorithm to the exact Hessians ∇ 2 f (xk ) have bounded condition numbers, that
is, the bound (3.39) holds for some value of C.
MODIFIED SYMMETRIC INDEFINITE FACTORIZATION
Another strategy for modifying an indefinite Hessian is to use a procedure based on
a symmetric indefinite factorization. Any symmetric matrix A, whether positive definite or
not, can be written as
P APT L BLT ,
(3.51)
where L is unit lower triangular, B is a block diagonal matrix with blocks of dimension 1
or 2, and P is a permutation matrix (see our discussion in Appendix A and also Golub and
3.4.
NEWTON’S METHOD
WITH
HESSIAN MODIFICATION
Van Loan [136, Section 4.4]). We mentioned earlier that attempting to compute the L DL T
factorization of an indefinite matrix (where D is a diagonal matrix) is inadvisable because
even if the factors L and D are well defined, they may contain entries that are larger than the
original elements of A, thus amplifying rounding errors that arise during the computation.
However, by using the block diagonal matrix B, which allows 2 × 2 blocks as well as 1 × 1
blocks on the diagonal, we can guarantee that the factorization (3.51) always exists and can
be computed by a numerically stable process.
❏ EXAMPLE 3.2
The matrix
⎡
0
⎢
⎢ 1
A⎢
⎢ 2
⎣
3
1
2
2
2
2
3
2
3
3
⎤
⎥
2 ⎥
⎥
3 ⎥
⎦
4
can be written in the form (3.51) with P [e1 , e4 , e3 , e2 ],
⎡
1
⎢
⎢ 0
⎢
1
L⎢
⎢
⎢ 9
⎣
2
9
0
0
1
2
3
1
3
0
1
0
0
⎤
⎥
0 ⎥
⎥
⎥,
0 ⎥
⎥
⎦
1
⎡
0
⎢
⎢ 3
⎢
B⎢
⎢ 0
⎢
⎣
0
3
0
0
4
0
7
9
5
9
0
5
9
10
9
0
0
⎤
⎥
⎥
⎥
⎥.
⎥
⎥
⎦
(3.52)
Note that both diagonal blocks in B are 2 × 2. Several algorithms for computing symmetric
indefinite factorizations are discussed in Section A.1 of Appendix A.
❐
The symmetric indefinite factorization allows us to determine the inertia of a matrix,
that is, the number of positive, zero, and negative eigenvalues. One can show that the inertia
of B equals the inertia of A. Moreover, the 2 × 2 blocks in B are always constructed to
have one positive and one negative eigenvalue. Thus the number of positive eigenvalues in
A equals the number of positive 1 × 1 blocks plus the number of 2 × 2 blocks.
As for the Cholesky factorization, an indefinite symmetric factorization algorithm
can be modified to ensure that the modified factors are the factors of a positive definite
matrix. The strategy is first to compute the factorization (3.51), as well as the spectral
decomposition B QQ T , which is inexpensive to compute because B is block diagonal
55
56
CHAPTER 3.
LINE SEARCH METHODS
(see Exercise 3.12). We then construct a modification matrix F such that
L(B + F)L T
is sufficiently positive definite. Motivated by the modified spectral decomposition (3.43),
we choose a parameter δ > 0 and define F to be
F Q diag(τi ) Q ,
T
τi 0,
λi ≥ δ,
δ − λi ,
λi < δ,
i 1, 2, . . . , n,
(3.53)
where λi are the eigenvalues of B. The matrix F is thus the modification of minimum
Frobenius norm that ensures that all eigenvalues of the modified matrix B + F are no less
than δ. This strategy therefore modifies the factorization (3.51) as follows:
P(A + E)P T L(B + F)L T ,
where E P T L F L T P.
(Note that E will not be diagonal, in general.) Hence, in contrast to the modified Cholesky
approach, this modification strategy changes the entire matrix A, not just its diagonal. The
aim of strategy (3.53) is that the modified matrix satisfies λmin (A + E) ≈ δ whenever the
original matrix A has λmin (A) < δ. It is not clear, however, whether it always comes close
to attaining this goal.
3.5
STEP-LENGTH SELECTION ALGORITHMS
We now consider techniques for finding a minimum of the one-dimensional function
φ(α) f (xk + αpk ),
(3.54)
or for simply finding a step length αk satisfying one of the termination conditions described
in Section 3.1. We assume that pk is a descent direction—that is, φ (0) < 0—so that our
search can be confined to positive values of α.
If f is a convex quadratic, f (x) 12 x T Qx − b T x, its one-dimensional minimizer
along the ray xk + αpk can be computed analytically and is given by
αk −
∇ f kT pk
.
pkT Qpk
(3.55)
For general nonlinear functions, it is necessary to use an iterative procedure. The line search
procedure deserves particular attention because it has a major impact on the robustness and
efficiency of all nonlinear optimization methods.
3.5.
STEP-LENGTH SELECTION ALGORITHMS
Line search procedures can be classified according to the type of derivative information
they use. Algorithms that use only function values can be inefficient since, to be theoretically
sound, they need to continue iterating until the search for the minimizer is narrowed down
to a small interval. In contrast, knowledge of gradient information allows us to determine
whether a suitable step length has been located, as stipulated, for example, by the Wolfe
conditions (3.6) or Goldstein conditions (3.11). Often, particularly when xk is close to the
solution, the very first choice of α satisfies these conditions, so the line search need not
be invoked at all. In the rest of this section, we discuss only algorithms that make use of
derivative information. More information on derivative-free procedures is given in the notes
at the end of this chapter.
All line search procedures require an initial estimate α0 and generate a sequence {αi }
that either terminates with a step length satisfying the conditions specified by the user (for
example, the Wolfe conditions) or determines that such a step length does not exist. Typical
procedures consist of two phases: a bracketing phase that finds an interval [ā, b̄] containing
acceptable step lengths, and a selection phase that zooms in to locate the final step length.
The selection phase usually reduces the bracketing interval during its search for the desired
step length and interpolates some of the function and derivative information gathered on
earlier steps to guess the location of the minimizer. We first discuss how to perform this
interpolation.
In the following discussion we let αk and αk−1 denote the step lengths used at iterations
k and k − 1 of the optimization algorithm, respectively. On the other hand, we denote the
trial step lengths generated during the line search by αi and αi−1 and also α j . We use α0 to
denote the initial guess.
INTERPOLATION
We begin by describing a line search procedure based on interpolation of known
function and derivative values of the function φ. This procedure can be viewed as an
enhancement of Algorithm 3.1. The aim is to find a value of α that satisfies the sufficient
decrease condition (3.6a), without being “too small.” Accordingly, the procedures here
generate a decreasing sequence of values αi such that each value αi is not too much smaller
than its predecessor αi−1 .
Note that we can write the sufficient decrease condition in the notation of (3.54) as
φ(αk ) ≤ φ(0) + c1 αk φ (0),
(3.56)
and that since the constant c1 is usually chosen to be small in practice (c1 10−4 , say), this
condition asks for little more than descent in f . We design the procedure to be “efficient”
in the sense that it computes the derivative ∇ f (x) as few times as possible.
Suppose that the initial guess α0 is given. If we have
φ(α0 ) ≤ φ(0) + c1 α0 φ (0),
57
58
CHAPTER 3.
LINE SEARCH METHODS
this step length satisfies the condition, and we terminate the search. Otherwise, we know that
the interval [0, α0 ] contains acceptable step lengths (see Figure 3.3). We form a quadratic
approximation φq (α) to φ by interpolating the three pieces of information available—φ(0),
φ (0), and φ(α0 )—to obtain
φq (α) φ(α0 ) − φ(0) − α0 φ (0)
α 2 + φ (0)α + φ(0).
α02
(3.57)
(Note that this function is constructed so that it satisfies the interpolation conditions
φq (0) φ(0), φq (0) φ (0), and φq (α0 ) φ(α0 ).) The new trial value α1 is defined as the
minimizer of this quadratic, that is, we obtain
α1 −
φ (0)α02
.
2 [φ(α0 ) − φ(0) − φ (0)α0 ]
(3.58)
If the sufficient decrease condition (3.56) is satisfied at α1 , we terminate the search. Otherwise, we construct a cubic function that interpolates the four pieces of information φ(0),
φ (0), φ(α0 ), and φ(α1 ), obtaining
φc (α) aα 3 + bα 2 + αφ (0) + φ(0),
where
a
b
1
2 2
α0 α1 (α1 − α0 )
α02
−α12
−α03
α13
φ(α1 ) − φ(0) − φ (0)α1
φ(α0 ) − φ(0) − φ (0)α0
.
By differentiating φc (x), we see that the minimizer α2 of φc lies in the interval [0, α1 ] and is
given by
α2 −b +
b2 − 3aφ (0)
.
3a
If necessary, this process is repeated, using a cubic interpolant of φ(0), φ (0) and the two
most recent values of φ, until an α that satisfies (3.56) is located. If any αi is either too
close to its predecessor αi−1 or else too much smaller than αi−1 , we reset αi αi−1 /2. This
safeguard procedure ensures that we make reasonable progress on each iteration and that
the final α is not too small.
The strategy just described assumes that derivative values are significantly more expensive to compute than function values. It is often possible, however, to compute the
directional derivative simultaneously with the function, at little additional cost; see Chapter 8. Accordingly, we can design an alternative strategy based on cubic interpolation of the
values of φ and φ at the two most recent values of α.
3.5.
STEP-LENGTH SELECTION ALGORITHMS
Cubic interpolation provides a good model for functions with significant changes of
curvature. Suppose we have an interval [ā, b̄] known to contain desirable step lengths, and
two previous step length estimates αi−1 and αi in this interval. We use a cubic function to
interpolate φ(αi−1 ), φ (αi−1 ), φ(αi ), and φ (αi ). (This cubic function always exists and is
unique; see, for example, Bulirsch and Stoer [41, p. 52].) The minimizer of this cubic in
[ā, b̄] is either at one of the endpoints or else in the interior, in which case it is given by
φ (αi ) + d2 − d1
,
(3.59)
αi+1 αi − (αi − αi−1 ) φ (αi ) − φ (αi−1 ) + 2d2
with
d1 φ (αi−1 ) + φ (αi ) − 3
φ(αi−1 ) − φ(αi )
,
αi−1 − αi
d2 sign(αi − αi−1 ) d12 − φ (αi−1 )φ (αi )
1/2
.
The interpolation process can be repeated by discarding the data at one of the step
lengths αi−1 or αi and replacing it by φ(αi+1 ) and φ (αi+1 ). The decision on which of αi−1
and αi should be kept and which discarded depends on the specific conditions used to
terminate the line search; we discuss this issue further below in the context of the Wolfe
conditions. Cubic interpolation is a powerful strategy, since it usually produces a quadratic
rate of convergence of the iteration (3.59) to the minimizing value of α.
INITIAL STEP LENGTH
For Newton and quasi-Newton methods, the step α0 1 should always be used as
the initial trial step length. This choice ensures that unit step lengths are taken whenever
they satisfy the termination conditions and allows the rapid rate-of-convergence properties
of these methods to take effect.
For methods that do not produce well scaled search directions, such as the steepest descent and conjugate gradient methods, it is important to use current information about the
problem and the algorithm to make the initial guess. A popular strategy is to assume that the
first-order change in the function at iterate xk will be the same as that obtained at the previT
pk−1 ,
ous step. In other words, we choose the initial guess α0 so that α0 ∇ f kT pk αk−1 ∇ f k−1
that is,
α0 αk−1
T
∇ f k−1
pk−1
∇ f kT pk
.
Another useful strategy is to interpolate a quadratic to the data f (xk−1 ), f (xk ), and
T
pk−1 and to define α0 to be its minimizer. This strategy yields
∇ f k−1
α0 2( f k − f k−1 )
.
φ (0)
(3.60)
59
60
CHAPTER 3.
LINE SEARCH METHODS
It can be shown that if xk → x ∗ superlinearly, then the ratio in this expression converges to
1. If we adjust the choice (3.60) by setting
α0 ← min(1, 1.01α0 ),
we find that the unit step length α0 1 will eventually always be tried and accepted, and the
superlinear convergence properties of Newton and quasi-Newton methods will be observed.
A LINE SEARCH ALGORITHM FOR THE WOLFE CONDITIONS
The Wolfe (or strong Wolfe) conditions are among the most widely applicable and
useful termination conditions. We now describe in some detail a one-dimensional search
procedure that is guaranteed to find a step length satisfying the strong Wolfe conditions (3.7)
for any parameters c1 and c2 satisfying 0 < c1 < c2 < 1. As before, we assume that p is a
descent direction and that f is bounded below along the direction p.
The algorithm has two stages. This first stage begins with a trial estimate α1 , and keeps
increasing it until it finds either an acceptable step length or an interval that brackets the
desired step lengths. In the latter case, the second stage is invoked by calling a function called
zoom (Algorithm 3.6, below), which successively decreases the size of the interval until an
acceptable step length is identified.
A formal specification of the line search algorithm follows. We refer to (3.7a) as the
sufficient decrease condition and to (3.7b) as the curvature condition. The parameter αmax
is a user-supplied bound on the maximum step length allowed. The line search algorithm
terminates with α∗ set to a step length that satisfies the strong Wolfe conditions.
Algorithm 3.5 (Line Search Algorithm).
Set α0 ← 0, choose αmax > 0 and α1 ∈ (0, αmax );
i ← 1;
repeat
Evaluate φ(αi );
if φ(αi ) > φ(0) + c1 αi φ (0) or [φ(αi ) ≥ φ(αi−1 ) and i > 1]
α∗ ←zoom(αi−1 , αi ) and stop;
Evaluate φ (αi );
if |φ (αi )| ≤ −c2 φ (0)
set α∗ ← αi and stop;
if φ (αi ) ≥ 0
set α∗ ←zoom(αi , αi−1 ) and stop;
Choose αi+1 ∈ (αi , αmax );
i ← i + 1;
end (repeat)
3.5.
STEP-LENGTH SELECTION ALGORITHMS
Note that the sequence of trial step lengths {αi } is monotonically increasing, but that
the order of the arguments supplied to the zoom function may vary. The procedure uses
the knowledge that the interval (αi−1 , αi ) contains step lengths satisfying the strong Wolfe
conditions if one of the following three conditions is satisfied:
(i) αi violates the sufficient decrease condition;
(ii) φ(αi ) ≥ φ(αi−1 );
(iii) φ (αi ) ≥ 0.
The last step of the algorithm performs extrapolation to find the next trial value αi+1 . To
implement this step we can use approaches like the interpolation procedures above, or
we can simply set αi+1 to some constant multiple of αi . Whichever strategy we use, it is
important that the successive steps increase quickly enough to reach the upper limit αmax in
a finite number of iterations.
We now specify the function zoom, which requires a little explanation. The order of
its input arguments is such that each call has the form zoom(αlo , αhi ), where
(a) the interval bounded by αlo and αhi contains step lengths that satisfy the strong Wolfe
conditions;
(b) αlo is, among all step lengths generated so far and satisfying the sufficient decrease
condition, the one giving the smallest function value; and
(c) αhi is chosen so that φ (αlo )(αhi − αlo ) < 0.
Each iteration of zoom generates an iterate α j between αlo and αhi , and then replaces one
of these endpoints by α j in such a way that the properties (a), (b), and (c) continue to hold.
Algorithm 3.6 (zoom).
repeat
Interpolate (using quadratic, cubic, or bisection) to find
a trial step length α j between αlo and αhi ;
Evaluate φ(α j );
if φ(α j ) > φ(0) + c1 α j φ (0) or φ(α j ) ≥ φ(αlo )
αhi ← α j ;
else
Evaluate φ (α j );
if |φ (α j )| ≤ −c2 φ (0)
Set α∗ ← α j and stop;
if φ (α j )(αhi − αlo ) ≥ 0
αhi ← αlo ;
αlo ← α j ;
end (repeat)
61
62
CHAPTER 3.
LINE SEARCH METHODS
If the new estimate α j happens to satisfy the strong Wolfe conditions, then zoom has served
its purpose of identifying such a point, so it terminates with α∗ α j . Otherwise, if α j
satisfies the sufficient decrease condition and has a lower function value than xlo , then we
set αlo ← α j to maintain condition (b). If this setting results in a violation of condition (c),
we remedy the situation by setting αhi to the old value of αlo . Readers should sketch some
graphs to see for themselves how zoom works!
As mentioned earlier, the interpolation step that determines α j should be safeguarded
to ensure that the new step length is not too close to the endpoints of the interval. Practical
line search algorithms also make use of the properties of the interpolating polynomials to
make educated guesses of where the next step length should lie; see [39, 216]. A problem
that can arise is that as the optimization algorithm approaches the solution, two consecutive
function values f (xk ) and f (xk−1 ) may be indistinguishable in finite-precision arithmetic.
Therefore, the line search must include a stopping test if it cannot attain a lower function
value after a certain number (typically, ten) of trial step lengths. Some procedures also
stop if the relative change in x is close to machine precision, or to some user-specified
threshold.
A line search algorithm that incorporates all these features is difficult to code. We
advocate the use of one of the several good software implementations available in the
public domain. See Dennis and Schnabel [92], Lemaréchal [189], Fletcher [101], Moré and
Thuente [216] (in particular), and Hager and Zhang [161].
One may ask how much more expensive it is to require the strong Wolfe conditions
instead of the regular Wolfe conditions. Our experience suggests that for a “loose” line
search (with parameters such as c1 10−4 and c2 0.9), both strategies require a similar
amount of work. The strong Wolfe conditions have the advantage that by decreasing c2 we
can directly control the quality of the search, by forcing the accepted value of α to lie closer
to a local minimum. This feature is important in steepest descent or nonlinear conjugate
gradient methods, and therefore a step selection routine that enforces the strong Wolfe
conditions has wide applicability.
NOTES AND REFERENCES
For an extensive discussion of line search termination conditions see Ortega and
Rheinboldt [230]. Akaike [2] presents a probabilistic analysis of the steepest descent method
with exact line searches on quadratic functions. He shows that when n > 2, the worst-case
bound (3.29) can be expected to hold for most starting points. The case n 2 can be
studied in closed form; see Bazaraa, Sherali, and Shetty [14]. Theorem 3.6 is due to Dennis
and Moré.
Some line search methods (see Goldfarb [132] and Moré and Sorensen [213]) compute
a direction of negative curvature, whenever it exists, to prevent the iteration from converging
to nonminimizing stationary points. A direction of negative curvature p− is one that satisfies
p−T ∇ 2 f (xk ) p− < 0. These algorithms generate a search direction by combining p− with the
steepest descent direction −∇ f k , often performing a curvilinear backtracking line search.
3.5.
STEP-LENGTH SELECTION ALGORITHMS
It is difficult to determine the relative contributions of the steepest descent and negative
curvature directions. Because of this fact, the approach fell out of favor after the introduction
of trust-region methods.
For a more thorough treatment of the modified Cholesky factorization see Gill,
Murray, and Wright [130] or Dennis and Schnabel [92]. A modified Cholesky factorization
based on Gershgorin disk estimates is described in Schnabel and Eskow [276]. The modified
indefinite factorization is from Cheng and Higham [58].
Another strategy for implementing a line search Newton method when the Hessian
contains negative eigenvalues is to compute a direction of negative curvature and use it to
define the search direction (see Moré and Sorensen [213] and Goldfarb [132]).
Derivative-free line search algorithms include golden section and Fibonacci search.
They share some of the features with the line search method given in this chapter. They
typically store three trial points that determine an interval containing a one-dimensional
minimizer. Golden section and Fibonacci differ in the way in which the trial step lengths are
generated; see, for example, [79, 39].
Our discussion of interpolation follows Dennis and Schnabel [92], and the algorithm
for finding a step length satisfying the strong Wolfe conditions can be found in Fletcher
[101].
✐
EXERCISES
✐
3.1 Program the steepest descent and Newton algorithms using the backtracking line
search, Algorithm 3.1. Use them to minimize the Rosenbrock function (2.22). Set the initial
step length α0 1 and print the step length used by each method at each iteration. First try
the initial point x0 (1.2, 1.2)T and then the more difficult starting point x0 (−1.2, 1)T .
✐
3.2 Show that if 0 < c2 < c1 < 1, there may be no step lengths that satisfy the Wolfe
conditions.
✐
3.3 Show that the one-dimensional minimizer of a strongly convex quadratic function
is given by (3.55).
✐
3.4 Show that the one-dimensional minimizer of a strongly convex quadratic function
always satisfies the Goldstein conditions (3.11).
✐
3.5 Prove that Bx ≥ x/B −1 for any nonsingular matrix B. Use this fact to
establish (3.19).
✐
3.6 Consider the steepest descent method with exact line searches applied to the
convex quadratic function (3.24). Using the properties given in this chapter, show that if the
initial point is such that x0 − x ∗ is parallel to an eigenvector of Q, then the steepest descent
method will find the solution in one step.
63
64
CHAPTER 3.
LINE SEARCH METHODS
✐
3.7 Prove the result (3.28) by working through the following steps. First, use (3.26)
to show that
xk − x ∗ 2Q − xk+1 − x ∗ 2Q 2αk ∇ f kT Q(xk − x ∗ ) − αk2 ∇ f kT Q∇ f k ,
where · Q is defined by (3.27). Second, use the fact that ∇ f k Q(xk − x ∗ ) to obtain
xk − x ∗ 2Q − xk+1 − x ∗ 2Q (∇ f kT ∇ f k )2
2(∇ f kT ∇ f k )2
−
(∇ f kT Q∇ f k ) (∇ f kT Q∇ f k )
and
xk − x ∗ 2Q ∇ f kT Q −1 ∇ f k .
✐
3.8 Let Q be a positive definite symmetric matrix. Prove that for any vector x, we
have
(x T
(x T x)2
4λn λ1
≥
,
Qx)(x T Q −1 x)
(λn + λ1 )2
where λn and λ1 are, respectively, the largest and smallest eigenvalues of Q. (This relation,
which is known as the Kantorovich inequality, can be used to deduce (3.29) from (3.28).)
✐
3.9 Program the BFGS algorithm using the line search algorithm described in this
chapter that implements the strong Wolfe conditions. Have the code verify that ykT sk is
always positive. Use it to minimize the Rosenbrock function using the starting points given
in Exercise 3.1.
✐
3.10 Compute the eigenvalues of the 2 diagonal blocks of (3.52) and verify that each
block has a positive and a negative eigenvalue. Then compute the eigenvalues of A and verify
that its inertia is the same as that of B.
✐
3.11 Describe the effect that the modified Cholesky factorization (3.50) would have
on the Hessian ∇ 2 f (xk ) diag(−2, 12, 4).
✐ 3.12 Consider a block diagonal matrix B with 1 × 1 and 2 × 2 blocks. Show that the
eigenvalues and eigenvectors of B can be obtained by computing the spectral decomposition
of each diagonal block separately.
✐
3.13 Show that the quadratic function that interpolates φ(0), φ (0), and φ(α0 ) is
given by (3.57). Then, make use of the fact that the sufficient decrease condition (3.6a) is
not satisfied at α0 to show that this quadratic has positive curvature and that the minimizer
satisfies
α1 <
α0
.
2(1 − c1 )
3.5.
STEP-LENGTH SELECTION ALGORITHMS
Since c1 is chosen to be quite small in practice, this inequality indicates that α1 cannot be
much greater than 12 (and may be smaller), which gives us an idea of the new step length.
✐
3.14 If φ(α0 ) is large, (3.58) shows that α1 can be quite small. Give an example of a
function and a step length α0 for which this situation arises. (Drastic changes to the estimate
of the step length are not desirable, since they indicate that the current interpolant does not
provide a good approximation to the function and that it should be modified before being
trusted to produce a good step length estimate. In practice, one imposes a lower bound—
typically, ρ 0.1—and defines the new step length as αi max(ραi−1 , α̂i ), where α̂i is the
minimizer of the interpolant.)
✐
3.15 Suppose that the sufficient decrease condition (3.6a) is not satisfied at the step
lengths α0 , and α1 , and consider the cubic interpolating φ(0), φ (0), φ(α0 ) and φ(α1 ).
By drawing graphs illustrating the two situations that can arise, show that the minimizer of the cubic lies in [0, α1 ]. Then show that if φ(0) < φ(α1 ), the minimizer is
less than 23 α1 .
65
This is pag
Printer: O
CHAPTER
4
Trust-Region
Methods
Line search methods and trust-region methods both generate steps with the help of a
quadratic model of the objective function, but they use this model in different ways. Line
search methods use it to generate a search direction, and then focus their efforts on finding
a suitable step length α along this direction. Trust-region methods define a region around
the current iterate within which they trust the model to be an adequate representation of
the objective function, and then choose the step to be the approximate minimizer of the
model in this region. In effect, they choose the direction and length of the step simultaneously. If a step is not acceptable, they reduce the size of the region and find a new
CHAPTER 4.
TRUST-REGION METHODS
minimizer. In general, the direction of the step changes whenever the size of the trust region
is altered.
The size of the trust region is critical to the effectiveness of each step. If the region is
too small, the algorithm misses an opportunity to take a substantial step that will move it
much closer to the minimizer of the objective function. If too large, the minimizer of the
model may be far from the minimizer of the objective function in the region, so we may have
to reduce the size of the region and try again. In practical algorithms, we choose the size of
the region according to the performance of the algorithm during previous iterations. If the
model is consistently reliable, producing good steps and accurately predicting the behavior
of the objective function along these steps, the size of the trust region may be increased to
allow longer, more ambitious, steps to be taken. A failed step is an indication that our model
is an inadequate representation of the objective function over the current trust region. After
such a step, we reduce the size of the region and try again.
Figure 4.1 illustrates the trust-region approach on a function f of two variables in
which the current point xk and the minimizer x ∗ lie at opposite ends of a curved valley.
The quadratic model function m k , whose elliptical contours are shown as dashed lines, is
constructed from function and derivative information at xk and possibly also on information
accumulated from previous iterations and steps. A line search method based on this model
searches along the step to the minimizer of m k (shown), but this direction will yield at most
a small reduction in f , even if the optimal steplength is used. The trust-region method
steps to the minimizer of m k within the dotted circle (shown), yielding a more significant
reduction in f and better progress toward the solution.
In this chapter, we will assume that the model function m k that is used at each
iterate xk is quadratic. Moreover, m k is based on the Taylor-series expansion of f around
Trust region
Line search direction
contours of mk
Trust region step
contours of f
Figure 4.1
Trust-region and line search steps.
67
68
CHAPTER 4.
TRUST-REGION METHODS
xk , which is
f (xk + p) f k + gkT p + 12 p T ∇ 2 f (xk + t p) p,
(4.1)
where f k f (xk ) and gk ∇ f (xk ), and t is some scalar in the interval (0, 1). By using an
approximation Bk to the Hessian in the second-order term, m k is defined as follows:
m k ( p) f k + gkT p + 12 p T Bk p,
(4.2)
where Bk is some symmetric matrix. The difference between m k ( p) and f (xk + p) is
O p2 , which is small when p is small.
When Bk is equal to the true Hessian ∇ 2 f (xk ), the approximation error in the model
function m k is O p3 , so this model is especially accurate when p is small. This choice
Bk ∇ 2 f (xk ) leads to the trust-region Newton method, and will be discussed further in
Section 4.4. In other sections of this chapter, we emphasize the generality of the trust-region
approach by assuming little about Bk except symmetry and uniform boundedness.
To obtain each step, we seek a solution of the subproblem
minn m k ( p) f k + gkT p + 12 p T Bk p
p∈IR
s.t. p ≤ k ,
(4.3)
where k > 0 is the trust-region radius. In most of our discussions, we define · to be
the Euclidean norm, so that the solution pk∗ of (4.3) is the minimizer of m k in the ball of
radius k . Thus, the trust-region approach requires us to solve a sequence of subproblems
(4.3) in which the objective function and constraint (which can be written as p T p ≤ 2k )
are both quadratic. When Bk is positive definite and Bk−1 gk ≤ k , the solution of (4.3) is
easy to identify—it is simply the unconstrained minimum pkB −Bk−1 gk of the quadratic
m k ( p). In this case, we call pkB the full step. The solution of (4.3) is not so obvious in other
cases, but it can usually be found without too much computational expense. In any case,
as described below, we need only an approximate solution to obtain convergence and good
practical behavior.
OUTLINE OF THE TRUST-REGION APPROACH
One of the key ingredients in a trust-region algorithm is the strategy for choosing the
trust-region radius k at each iteration. We base this choice on the agreement between the
model function m k and the objective function f at previous iterations. Given a step pk we
define the ratio
ρk f (xk ) − f (xk + pk )
;
m k (0) − m k ( pk )
(4.4)
the numerator is called the actual reduction, and the denominator is the predicted reduction
(that is, the reduction in f predicted by the model function). Note that since the step pk
CHAPTER 4.
TRUST-REGION METHODS
is obtained by minimizing the model m k over a region that includes p 0, the predicted
reduction will always be nonnegative. Hence, if ρk is negative, the new objective value
f (xk + pk ) is greater than the current value f (xk ), so the step must be rejected. On the
other hand, if ρk is close to 1, there is good agreement between the model m k and the
function f over this step, so it is safe to expand the trust region for the next iteration. If ρk
is positive but significantly smaller than 1, we do not alter the trust region, but if it is close
to zero or negative, we shrink the trust region by reducing k at the next iteration.
The following algorithm describes the process.
Algorithm 4.1 (Trust Region).
ˆ > 0, 0 ∈ (0, ),
ˆ and η ∈ 0, 1 :
Given 4
for k 0, 1, 2, . . .
Obtain pk by (approximately) solving (4.3);
Evaluate ρk from (4.4);
if ρk < 14
k+1 14 k
else
if ρk > 34 and pk k
ˆ
k+1 min(2k , )
else
k+1 k ;
if ρk > η
xk+1 xk + pk
else
xk+1 xk ;
end (for).
ˆ is an overall bound on the step lengths. Note that the radius is increased only if pk Here actually reaches the boundary of the trust region. If the step stays strictly inside the region,
we infer that the current value of k is not interfering with the progress of the algorithm,
so we leave its value unchanged for the next iteration.
To turn Algorithm 4.1 into a practical algorithm, we need to focus on solving the
trust-region subproblem (4.3). In discussing this matter, we sometimes drop the iteration
subscript k and restate the problem (4.3) as follows:
def
minn m( p) f + g T p + 12 p T Bp
p∈IR
s.t. p ≤ .
(4.5)
A first step to characterizing exact solutions of (4.5) is given by the following theorem
(due to Moré and Sorensen [214]), which shows that the solution p ∗ of (4.5) satisfies
(B + λI ) p ∗ −g
for some λ ≥ 0.
(4.6)
69
70
CHAPTER 4.
TRUST-REGION METHODS
Theorem 4.1.
The vector p ∗ is a global solution of the trust-region problem
minn m( p) f + g T p + 12 p T Bp,
p∈IR
s.t. p ≤ ,
(4.7)
if and only if p ∗ is feasible and there is a scalar λ ≥ 0 such that the following conditions are
satisfied:
(B + λI ) p ∗ −g,
(4.8a)
∗
λ( − || p ||) 0,
(4.8b)
(B + λI )
(4.8c)
is positive semidefinite.
We delay the proof of this result until Section 4.3, and instead discuss just its key
features here with the help of Figure 4.2. The condition (4.8b) is a complementarity condition
that states that at least one of the nonnegative quantities λ and ( − p ∗ ) must be zero.
Hence, when the solution lies strictly inside the trust region (as it does when 1 in
Figure 4.2), we must have λ 0 and so Bp ∗ −g with B positive semidefinite, from (4.8a)
and (4.8c), respectively. In the other cases 2 and 3 , we have p ∗ , and
so λ is allowed to take a positive value. Note from (4.8a) that
λp ∗ −Bp ∗ − g −∇m( p ∗ ).
∆1
∆2
∆3
contours of m
p*1
Figure 4.2
p*2
p*3
Solution of trust-region subproblem for different radii 1 , 2 , 3 .
4.1.
ALGORITHMS BASED
ON THE
CAUCHY POINT
Thus, when λ > 0, the solution p ∗ is collinear with the negative gradient of m and normal
to its contours. These properties can be seen in Figure 4.2.
In Section 4.1, we describe two strategies for finding approximate solutions of the
subproblem (4.3), which achieve at least as much reduction in m k as the reduction achieved
by the so-called Cauchy point. This point is simply the minimizer of m k along the steepest
descent direction −gk . subject to the trust-region bound. The first approximate strategy is
the dogleg method, which is appropriate when the model Hessian Bk is positive definite. The
second strategy, known as two-dimensional subspace minimization, can be applied when Bk
is indefinite, though it requires an estimate of the most negative eigenvalue of this matrix.
A third strategy, described in Section 7.1, uses an approach based on the conjugate gradient
method to minimize m k , and can therefore be applied when B is large and sparse.
Section 4.3 is devoted to a strategy in which an iterative method is used to identify the
value of λ for which (4.6) is satisfied by the solution of the subproblem. We prove global
convergence results in Section 4.2. Section 4.4 discusses the trust-region Newton method, in
which the Hessian Bk of the model function is equal to the Hessian ∇ 2 f (xk ) of the objective
function. The key result of this section is that, when the trust-region Newton algorithm converges to a point x ∗ satisfying second-order sufficient conditions, it converges superlinearly.
4.1
ALGORITHMS BASED ON THE CAUCHY POINT
THE CAUCHY POINT
As we saw in Chapter 3, line search methods can be globally convergent even when the
optimal step length is not used at each iteration. In fact, the step length αk need only satisfy
fairly loose criteria. A similar situation applies in trust-region methods. Although in principle
we seek the optimal solution of the subproblem (4.3), it is enough for purposes of global
convergence to find an approximate solution pk that lies within the trust region and gives a
sufficient reduction in the model. The sufficient reduction can be quantified in terms of the
Cauchy point, which we denote by pkC and define in terms of the following simple procedure.
Algorithm 4.2 (Cauchy Point Calculation).
Find the vector pkS that solves a linear version of (4.3), that is,
pkS arg minn f k + gkT p
p∈IR
s.t. p ≤ k ;
(4.9)
Calculate the scalar τk > 0 that minimizes m k (τ pkS ) subject to
satisfying the trust-region bound, that is,
τk arg min m k (τ pkS )
τ ≥0
Set pkC τk pkS .
s.t. τ pkS ≤ k ;
(4.10)
71
72
CHAPTER 4.
TRUST-REGION METHODS
It is easy to write down a closed-form definition of the Cauchy point. For a start, the
solution of (4.9) is simply
pkS −
k
gk .
gk To obtain τk explicitly, we consider the cases of gkT Bk gk ≤ 0 and gkT Bk gk > 0 separately. For
the former case, the function m k (τ pkS ) decreases monotonically with τ whenever gk 0,
so τk is simply the largest value that satisfies the trust-region bound, namely, τk 1. For
the case gkT Bk gk > 0, m k (τ pkS ) is a convex quadratic in τ , so τk is either the unconstrained
minimizer of this quadratic, gk 3 /(k gkT Bk gk ), or the boundary value 1, whichever comes
first. In summary, we have
pkC −τk
k
gk ,
gk (4.11)
where
τk 1
min gk 3 /(k gkT Bk gk ), 1
if gkT Bk gk ≤ 0;
otherwise.
(4.12)
Figure 4.3 illustrates the Cauchy point for a subproblem in which Bk is positive
definite. In this example, pkC lies strictly inside the trust region.
The Cauchy step pkC is inexpensive to calculate—no matrix factorizations are
required—and is of crucial importance in deciding if an approximate solution of the
trust-region subproblem is acceptable. Specifically, a trust-region method will be globally
Trust region
contours of mk
pC
k
—gk
Figure 4.3
The Cauchy point.
4.1.
ALGORITHMS BASED
ON THE
CAUCHY POINT
convergent if its steps pk give a reduction in the model m k that is at least some fixed positive
multiple of the decrease attained by the Cauchy step.
IMPROVING ON THE CAUCHY POINT
Since the Cauchy point pkC provides sufficient reduction in the model function m k
to yield global convergence, and since the cost of calculating it is so small, why should
we look any further for a better approximate solution of (4.3)? The reason is that by
always taking the Cauchy point as our step, we are simply implementing the steepest
descent method with a particular choice of step length. As we have seen in Chapter 3, steepest descent performs poorly even if an optimal step length is used at each
iteration.
The Cauchy point does not depend very strongly on the matrix Bk , which is used only
in the calculation of the step length. Rapid convergence can be expected only if Bk plays a
role in determining the direction of the step as well as its length, and if Bk contains valid
curvature information about the function.
A number of trust-region algorithms compute the Cauchy point and then try to
improve on it. The improvement strategy is often designed so that the full step pkB −Bk−1 gk
is chosen whenever Bk is positive definite and pkB ≤ k . When Bk is the exact Hessian
∇ 2 f (xk ) or a quasi-Newton approximation, this strategy can be expected to yield superlinear
convergence.
We now consider three methods for finding approximate solutions to (4.3) that have
the features just described. Throughout this section we will be focusing on the internal
workings of a single iteration, so we simplify the notation by dropping the subscript “k”
from the quantities k , pk , m k , and gk and refer to the formulation (4.5) of the subproblem.
In this section, we denote the solution of (4.5) by p ∗ (), to emphasize the dependence
on .
THE DOGLEG METHOD
The first approach we discuss goes by the descriptive title of the dogleg method. It can
be used when B is positive definite.
To motivate this method, we start by examining the effect of the trust-region radius on the solution p ∗ () of the subproblem (4.5). When B is positive definite, we have already
noted that the unconstrained minimizer of m is p B −B −1 g. When this point is feasible
for (4.5), it is obviously a solution, so we have
p ∗ () p B ,
when ≥ p B .
(4.13)
When is small relative to p B , the restriction p ≤ ensures that the quadratic term in
m has little effect on the solution of (4.5). For such , we can get an approximation to p()
73
74
CHAPTER 4.
TRUST-REGION METHODS
Trust region
Optimal trajectory p (∆ )
pB ( full step )
pU ( unconstrained min along —g)
—g
dogleg path
Figure 4.4
Exact trajectory and dogleg approximation.
by simply omitting the quadratic term from (4.5) and writing
p ∗ () ≈ −
g
,
g
when is small.
(4.14)
For intermediate values of , the solution p ∗ () typically follows a curved trajectory like
the one in Figure 4.4.
The dogleg method finds an approximate solution by replacing the curved trajectory
for p ∗ () with a path consisting of two line segments. The first line segment runs from the
origin to the minimizer of m along the steepest descent direction, which is
pU −
gT g
g,
g T Bg
(4.15)
while the second line segment runs from p U to p B (see Figure 4.4). Formally, we denote this
trajectory by p̃(τ ) for τ ∈ [0, 2], where
p̃(τ ) τ pU ,
0 ≤ τ ≤ 1,
p + (τ − 1)( p − p ),
U
B
U
1 ≤ τ ≤ 2.
(4.16)
The dogleg method chooses p to minimize the model m along this path, subject to
the trust-region bound. The following lemma shows that the minimum along the dogleg
path can be found easily.
4.1.
ALGORITHMS BASED
ON THE
CAUCHY POINT
Lemma 4.2.
Let B be positive definite. Then
(i) p̃(τ ) is an increasing function of τ , and
(ii) m( p̃(τ )) is a decreasing function of τ .
PROOF. It is easy to show that (i) and (ii) both hold for τ ∈ [0, 1], so we restrict our
attention to the case of τ ∈ [1, 2]. For (i), define h(α) by
h(α) 12 p̃(1 + α)2
12 p U + α( p B − p U )2
12 p U 2 + α( p U )T ( p B − p U ) + 12 α 2 p B − p U 2 .
Our result is proved if we can show that h (α) ≥ 0 for α ∈ (0, 1). Now,
h (α) −( p U )T ( p U − p B ) + α p U − p B 2
≥ −( p U )T ( p U − p B )
gT g
gT g T
−1
g − T
g+B g
T
g Bg
g Bg
−1
(g T g)2
T gB g
g g T
1− T
g Bg
(g Bg)(g T B −1 g)
≥ 0,
where the final inequality is a consequence of the Cauchy-Schwarz inequality. (We leave the
details as an exercise.)
For (ii), we define ĥ(α) m( p̃(1 + α)) and show that ĥ (α) ≤ 0 for α ∈ (0, 1).
Substitution of (4.16) into (4.5) and differentiation with respect to the argument leads to
ĥ (α) ( p B − p U )T (g + B p U ) + α( p B − p U )T B( p B − p U )
≤ ( p B − p U )T (g + B p U + B( p B − p U ))
( p B − p U )T (g + B p B ) 0,
giving the result.
It follows from this lemma that the path p̃(τ ) intersects the trust-region boundary
p at exactly one point if p B ≥ , and nowhere otherwise. Since m is decreasing
along the path, the chosen value of p will be at p B if p B ≤ , otherwise at the point of
intersection of the dogleg and the trust-region boundary. In the latter case, we compute the
appropriate value of τ by solving the following scalar quadratic equation:
p U + (τ − 1)( p B − p U )2 2 .
75
76
CHAPTER 4.
TRUST-REGION METHODS
Consider now the case in which the exact Hessian ∇ 2 f (xk ) is available for use in the
model problem (4.5). When ∇ 2 f (xk ) is positive definite, we can simply set B ∇ 2 f (xk )
(that is, p B (∇ 2 f (xk ))−1 gk ) and apply the procedure above to find the Newton-dogleg
step. Otherwise, we can define p B by choosing B to be one of the positive definite modified
Hessians described in Section 3.4, then proceed as above to find the dogleg step. Near
a solution satisfying second-order sufficient conditions (see Theorem 2.4), p B will be set
to the usual Newton step, allowing the possibility of rapid local convergence of Newton’s
method (see Section 4.4).
The use of a modified Hessian in the Newton-dogleg method is not completely
satisfying from an intuitive viewpoint, however. A modified factorization perturbs the
diagonals of ∇ 2 f (xk ) in a somewhat arbitrary manner, and the benefits of the trust-region
approach may not be realized. In fact, the modification introduced during the factorization
of the Hessian is redundant in some sense because the trust-region strategy introduces its
own modification. As we show in Section 4.3, the exact solution of the trust-region problem
(4.3) with Bk ∇ 2 f (xk ) is (∇ 2 f (xk ) + λI )−1 gk , where λ is chosen large enough to make
(∇ 2 f (xk ) + λI ) positive definite, and its value depends on the trust-region radius k . We
conclude that the Newton-dogleg method is most appropriate when the objective function
is convex (that is, ∇ 2 f (xk ) is always positive semidefinite). The techniques described below
may be more suitable for the general case.
The dogleg strategy can be adapted to handle indefinite matrices B, but there is not
much point in doing so because the full step p B is not the unconstrained minimizer of m
in this case. Instead, we now describe another strategy, which aims to include directions of
negative curvature (that is, directions d for which d T Bd < 0) in the space of candidate
trust-region steps.
TWO-DIMENSIONAL SUBSPACE MINIMIZATION
When B is positive definite, the dogleg method strategy can be made slightly more
sophisticated by widening the search for p to the entire two-dimensional subspace spanned
by p U and p B (equivalently, g and −B −1 g). The subproblem (4.5) is replaced by
min m( p) f + g T p + 12 p T Bp
p
s.t. p ≤ , p ∈ span[g, B −1 g].
(4.17)
This is a problem in two variables that is computationally inexpensive to solve. (After some
algebraic manipulation it can be reduced to finding the roots of a fourth degree polynomial.)
Clearly, the Cauchy point p C is feasible for (4.17), so the optimal solution of this subproblem
yields at least as much reduction in m as the Cauchy point, resulting in global convergence
of the algorithm. The two-dimensional subspace minimization strategy is obviously an
extension of the dogleg method as well, since the entire dogleg path lies in span[g, B −1 g].
This strategy can be modified to handle the case of indefinite B in a way that is intuitive,
practical, and theoretically sound. We mention just the salient points of the handling of the
4.2.
GLOBAL CONVERGENCE
indefiniteness here, and refer the reader to papers by Byrd, Schnabel, and Schultz (see [54]
and [279]) for details. When B has negative eigenvalues, the two-dimensional subspace in
(4.17) is changed to
span[g, (B + α I )−1 g],
for some α ∈ (−λ1 , −2λ1 ],
(4.18)
where λ1 denotes the most negative eigenvalue of B. (This choice of α ensures that B + α I is
positive definite, and the flexibility in the choice of α allows us to use a numerical procedure
such as the Lanczos method to compute it.) When (B + α I )−1 g ≤ , we discard the
subspace search of (4.17), (4.18) and instead define the step to be
p −(B + α I )−1 g + v,
(4.19)
where v is a vector that satisfies v T (B + α I )−1 g ≤ 0. (This condition ensures that p ≥
(B + α I )−1 g.) When B has zero eigenvalues but no negative eigenvalues, we define the
step to be the Cauchy point p p C .
When the exact Hessian is available, we can set B ∇ 2 f (xk ), and note that B −1 g is
the Newton step. Hence, when the Hessian is positive definite at the solution x ∗ and when
xk is close to x ∗ and is sufficiently large, the subspace minimization problem (4.17) will
be solved by the Newton step.
The reduction in model function m achieved by the two-dimensional subspace minimization strategy often is close to the reduction achieved by the exact solution of (4.5).
Most of the computational effort lies in a single factorization of B or B + α I (estimation of
α and solution of (4.17) are less significant), while strategies that find nearly exact solutions
of (4.5) typically require two or three such factorizations (see Section 4.3).
4.2
GLOBAL CONVERGENCE
REDUCTION OBTAINED BY THE CAUCHY POINT
In the preceding discussion of algorithms for approximately solving the trust-region
subproblem, we have repeatedly emphasized that global convergence depends on the approximate solution obtaining at least as much decrease in the model function m as the
Cauchy point. (In fact, a fixed positive fraction of the Cauchy decrease suffices.) We start
the global convergence analysis by obtaining an estimate of the decrease in m achieved by
the Cauchy point. We then use this estimate to prove that the sequence of gradients {gk }
generated by Algorithm 4.1 has an accumulation point at zero, and in fact converges to zero
when η is strictly positive.
Our first main result is that the dogleg and two-dimensional subspace minimization
algorithms and Steihaug’s algorithm (Algorithm 7.2) produce approximate solutions pk of
the subproblem (4.3) that satisfy the following estimate of decrease in the model function:
gk ,
m k (0) − m k ( pk ) ≥ c1 gk min k ,
Bk (4.20)
77
78
CHAPTER 4.
TRUST-REGION METHODS
for some constant c1 ∈ (0, 1]. The usefulness of this estimate will become clear in the
following two sections. For now, we note that when k is the minimum value in (4.20), the
condition is slightly reminiscent of the first Wolfe condition: The desired reduction in the
model is proportional to the gradient and the size of the step.
We show now that the Cauchy point pkC satisfies (4.20), with c1 12 .
Lemma 4.3.
The Cauchy point pkC satisfies (4.20) with c1 12 , that is,
gk m k (0) − m k ( pkC ) ≥ 12 gk min k ,
.
Bk (4.21)
PROOF. For simplicity, we drop the iteration index k in the proof.
We consider first the case g T Bg ≤ 0. Here, we have
m( p C ) − m(0) m(−g/g) − f
2 T
−
g2 + 12
g Bg
g
g2
≤ −g
g
≤ −g min ,
,
B
and so (4.21) certainly holds.
For the next case, consider g T Bg > 0 and
g3
≤ 1.
g T Bg
From (4.12), we have τ g3 / g T Bg , and so from (4.11) it follows that
(4.22)
g4
g4
1 T
+
g
Bg
2
g T Bg
(g T Bg)2
4
g
− 12 T
g Bg
g4
≤ − 12
Bg2
g2
− 12
B
g
≤ − 12 g min ,
,
B
m( p C ) − m(0) −
so (4.21) holds here too.
In the remaining case, (4.22) does not hold, and therefore
g T Bg <
g3
.
(4.23)
4.2.
GLOBAL CONVERGENCE
From (4.12), we have τ 1, and using this fact together with (4.23), we obtain
1 2 T
g2 +
g Bg
g
2 g2
1 2 g3
≤ −g +
2 g2 1
− 2 g
g
1
,
≤ − 2 g min ,
B
m( p C ) − m(0) −
yielding the desired result (4.21) once again.
To satisfy (4.20), our approximate solution pk has only to achieve a reduction that is
at least some fixed fraction c2 of the reduction achieved by the Cauchy point. We state the
observation formally as a theorem.
Theorem 4.4.
Let pk be any vector such that pk ≤ k and m k (0)−m k ( pk ) ≥ c2 m k (0) − m k ( pkC ) .
Then pk satisfies (4.20) with c1 c2 /2. In particular, if pk is the exact solution pk∗ of (4.3),
then it satisfies (4.20) with c1 12 .
PROOF. Since pk ≤ k , we have from Lemma 4.3 that
1
gk C
m k (0) − m k ( pk ) ≥ c2 m k (0) − m k ( pk ) ≥ 2 c2 gk min k ,
,
Bk giving the result.
Note that the dogleg and two-dimensional subspace minimization algorithms both
satisfy (4.20) with c1 12 , because they all produce approximate solutions pk for which
m k ( pk ) ≤ m k ( pkC ).
CONVERGENCE TO STATIONARY POINTS
Global convergence results for trust-region methods come in two varieties, depending
on whether we set the parameter η in Algorithm 4.1 to zero or to some small positive value.
When η 0 (that is, the step is taken whenever it produces a lower value of f ), we can
show that the sequence of gradients {gk } has a limit point at zero. For the more stringent
acceptance test with η > 0, which requires the actual decrease in f to be at least some small
fraction of the predicted decrease, we have the stronger result that gk → 0.
In this section we prove the global convergence results for both cases. We assume
throughout that the approximate Hessians Bk are uniformly bounded in norm, and that f
79
80
CHAPTER 4.
TRUST-REGION METHODS
is bounded below on the level set
def
S {x | f (x) ≤ f (x0 )}.
(4.24)
For later reference, we define an open neighborhood of this set by
def
S(R0 ) {x | x − y < R0 for some y ∈ S},
where R0 is a positive constant.
To allow our results to be applied more generally, we also allow the length of the
approximate solution pk of (4.3) to exceed the trust-region bound, provided that it stays
within some fixed multiple of the bound; that is,
pk ≤ γ k ,
for some constant γ ≥ 1.
(4.25)
The first result deals with the case η 0.
Theorem 4.5.
Let η 0 in Algorithm 4.1. Suppose that Bk ≤ β for some constant β, that f is
bounded below on the level set S defined by (4.24) and Lipschitz continuously differentiable in
the neighborhood S(R0 ) for some R0 > 0, and that all approximate solutions of (4.3) satisfy
the inequalities (4.20) and (4.25), for some positive constants c1 and γ . We then have
lim inf gk 0.
k→∞
(4.26)
PROOF. By performing some technical manipulation with the ratio ρk from (4.4), we obtain
( f (xk ) − f (xk + pk )) − (m k (0) − m k ( pk )) |ρk − 1| m k (0) − m k ( pk )
m k ( pk ) − f (xk + pk ) .
m k (0) − m k ( pk ) Since from Taylor’s theorem (Theorem 2.1) we have that
1
f (xk + pk ) f (xk ) + g(xk )T pk +
[g(xk + t pk ) − g(xk )]T pk dt,
0
for some t ∈ (0, 1), it follows from the definition (4.2) of m k that
1
1 T
T
[g(xk + t pk ) − g(xk )] pk dt |m k ( pk ) − f (xk + pk )| 2 pk Bk pk −
0
≤ (β/2) pk 2 + β1 pk 2 ,
(4.27)
4.2.
GLOBAL CONVERGENCE
where we have used β1 to denote the Lipschitz constant for g on the set S(R0 ), and assumed
that pk ≤ R0 to ensure that xk and xk + t pk both lie in the set S(R0 ).
Suppose for contradiction that there is > 0 and a positive index K such that
gk ≥ ,
for all k ≥ K .
(4.28)
From (4.20), we have for k ≥ K that
gk ≥ c1 min k ,
.
m k (0) − m k ( pk ) ≥ c1 gk min k ,
Bk β
(4.29)
Using (4.29), (4.27), and the bound (4.25), we have
|ρk − 1| ≤
γ 2 2k (β/2 + β1 )
.
c1 min(k , /β)
(4.30)
We now derive a bound on the right-hand-side that holds for all sufficiently small values of
¯ where ¯ is defined as follows:
k , that is, for all k ≤ ,
¯ min
c1
R0
1
,
2
2 γ (β/2 + β1 ) γ
.
(4.31)
The R0 /γ term in this definition ensures that the bound (4.27) is valid (because pk ≤
¯ ≤ R0 ). Note that since c1 ≤ 1 and γ ≥ 1, we have ¯ ≤ /β. The latter
γ k ≤ γ ¯
condition implies that for all k ∈ [0, ], we have min(k , /β) k , so from (4.30) and
(4.31), we have
|ρk − 1| ≤
¯
γ 2 2k (β/2 + β1 )
+ β1 )
γ 2 (β/2
1
γ 2 k (β/2 + β1 )
≤
≤ .
c1 k
c1
c1
2
Therefore, ρk > 14 , and so by the workings of Algorithm 4.1, we have k+1 ≥ k whenever
¯ It follows that reduction of k by a factor of 1 can occur
k falls below the threshold .
4
in our algorithm only if
¯
k ≥ ,
and therefore we conclude that
¯
k ≥ min K , /4
for all k ≥ K .
Suppose now that there is an infinite subsequence K such that ρk ≥
(4.32)
1
4
for k ∈ K. For
81
82
CHAPTER 4.
TRUST-REGION METHODS
k ∈ K and k ≥ K , we have from (4.29) that
f (xk ) − f (xk+1 ) f (xk ) − f (xk + pk )
≥
≥
1
(0) − m k ( pk )]
4 [m k
1
c min(k , /β).
4 1
Since f is bounded below, it follows from this inequality that
lim
k∈K, k→∞
k 0,
contradicting (4.32). Hence no such infinite subsequence K can exist, and we must have
ρk < 14 for all k sufficiently large. In this case, k will eventually be multiplied by 14 at every
iteration, and we have limk→∞ k 0, which again contradicts (4.32). Hence, our original
assertion (4.28) must be false, giving (4.26).
Our second global convergence result, for the case η > 0, borrows much of the analysis
from the proof above. Our approach here follows that of Schultz, Schnabel, and Byrd [279].
Theorem 4.6.
Let η ∈ 0, 14 in Algorithm 4.1. Suppose that Bk ≤ β for some constant β, that f is
bounded below on the level set S (4.24) and Lipschitz continuously differentiable in S(R0 ) for
some R0 > 0, and that all approximate solutions pk of (4.3) satisfy the inequalities (4.20) and
(4.25) for some positive constants c1 and γ . We then have
lim gk 0.
k→∞
(4.33)
PROOF. We consider a particular positive index m with gm 0. Using β1 again to denote
the Lipschitz constant for g on the set S(R0 ), we have
g(x) − gm ≤ β1 x − xm ,
for all x ∈ S(R0 ). We now define the scalars and R to satisfy
1
2 gm ,
R min
, R0 .
β1
Note that the ball
B(xm , R) {x | x − xm ≤ R}
is contained in S(R0 ), so Lipschitz continuity of g holds inside B(xm , R). We have
x ∈ B(xm , R) ⇒ g(x) ≥ gm − g(x) − gm ≥ 12 gm .
If the entire sequence {xk }k≥m stays inside the ball B(xm , R), we would have gk ≥
>0
4.3.
ITERATIVE SOLUTION
OF THE
SUBPROBLEM
for all k ≥ m. The reasoning in the proof of Theorem 4.5 can be used to show that this
scenario does not occur. Therefore, the sequence {xk }k≥m eventually leaves B(xm , R).
Let the index l ≥ m be such that xl+1 is the first iterate after xm outside B(xm , R).
Since gk ≥ for k m, m + 1, . . . , l, we can use (4.29) to write
f (xm ) − f (xl+1 ) l
f (xk ) − f (xk+1 )
km
l
η[m k (0) − m k ( pk )]
l
,
≥
ηc1 min k ,
km,xk xk+1
β
≥
km,xk xk+1
where we have limited the sum to the iterations k for which xk xk+1 , that is, those iterations
on which a step was actually taken. If k ≤ /β for all k m, m + 1, . . . , l, we have
f (xm ) − f (xl+1 ) ≥ ηc1
l
k ≥ ηc1 R ηc1 min
km,xk xk+1
β1
, R0 .
(4.34)
Otherwise, we have k > /β for some k m, m + 1, . . . , l, and so
f (xm ) − f (xl+1 ) ≥ ηc1
β
.
(4.35)
Since the sequence { f (xk )}∞
k0 is decreasing and bounded below, we have that
f (xk ) ↓ f ∗
(4.36)
for some f ∗ > −∞. Therefore, using (4.34) and (4.35), we can write
f (xm ) − f ∗ ≥ f (xm ) − f (xl+1 )
, , R0
≥ ηc1 min
β β1
gm gm 1
,
, R0 > 0.
ηc1 gm min
2
2β
2β1
Since f (xm ) − f ∗ ↓ 0, we must have gm → 0, giving the result.
4.3
ITERATIVE SOLUTION OF THE SUBPROBLEM
In this section, we describe a technique that uses the characterization (4.6) of the subproblem solution, applying Newton’s method to find the value of λ which matches the given
83
84
CHAPTER 4.
TRUST-REGION METHODS
trust-region radius in (4.5). We also prove the key result Theorem 4.1 concerning the
characterization of solutions of (4.3).
The methods of Section 4.1 make no serious attempt to find the exact solution of
the subproblem (4.5). They do, however, make some use of the information in the model
Hessian Bk , and they have advantages of reasonable implementation cost and nice global
convergence properties.
When the problem is relatively small (that is, n is not too large), it may be worthwhile
to exploit the model more fully by looking for a closer approximation to the solution of the
subproblem. In this section, we describe an approach for finding a good approximation at the
cost of a few factorizations of the matrix B (typically three factorization), as compared with
a single factorization for the dogleg and two-dimensional subspace minimization methods.
This approach is based on the characterization of the exact solution given in Theorem 4.1,
together with an ingenious application of Newton’s method in one variable. Essentially, the
algorithm tries to identify the value of λ for which (4.6) is satisfied by the solution of (4.5).
The characterization of Theorem 4.1 suggests an algorithm for finding the solution p
of (4.7). Either λ 0 satisfies (4.8a) and (4.8c) with p ≤ , or else we define
p(λ) −(B + λI )−1 g
for λ sufficiently large that B + λI is positive definite and seek a value λ > 0 such that
p(λ) .
(4.37)
This problem is a one-dimensional root-finding problem in the variable λ.
To see that a value of λ with all the desired properties exists, we appeal to the eigendecomposition of B and use it to study the properties of p(λ). Since B is symmetric, there
is an orthogonal matrix Q and a diagonal matrix such that B QQ T , where
diag(λ1 , λ2 , . . . , λn ),
and λ1 ≤ λ2 ≤ · · · ≤ λn are the eigenvalues of B; see (A.16). Clearly, B + λI Q( +
λI )Q T , and for λ λ j , we have
p(λ) −Q( + λI )−1 Q T g −
n
q Tj g
j1
λj + λ
qj,
(4.38)
where q j denotes the jth column of Q. Therefore, by orthonormality of q1 , q2 , . . . , qn , we
have
p(λ)2 n
2
q Tj g
j1
(λ j + λ)2
.
(4.39)
4.3.
ITERATIVE SOLUTION
2
1
OF THE
SUBPROBLEM
|| p ||
*
3
Figure 4.5
p(λ) as a function of λ.
This expression tells us a lot about p(λ). If λ > −λ1 , we have λ j + λ > 0 for all
j 1, 2, . . . , n, and so p(λ) is a continuous, nonincreasing function of λ on the interval
(−λ1 , ∞). In fact, we have that
lim p(λ) 0.
λ→∞
(4.40)
Moreover, we have when q Tj g 0 that
lim p(λ) ∞.
λ→−λ j
(4.41)
Figure 4.5 plots p(λ) against λ in a case in whcih q1T g, q2T g, and q3T g are all nonzero.
Note that the properties (4.40) and (4.41) hold and that p(λ) is a nonincreasing function
of λ on (−λ1 , ∞). In particular, as is always the case when q1T g 0, that there is a unique
value λ∗ ∈ (−λ1 , ∞) such that p(λ∗ ) . (There may be other, smaller values of λ for
which p(λ) , but these will fail to satisfy (4.8c).)
We now sketch a procedure for identifying the λ∗ ∈ (−λ1 , ∞) for which p(λ∗ ) ,
which works when q1T g 0. (We discuss the case of q1T g 0 later.) First, note that when B
positive definite and B −1 g ≤ , the value λ 0 satisfies (4.8), so the procedure can be
terminated immediately with λ∗ 0. Otherwise, we could use the root-finding Newton’s
method (see the Appendix) to find the value of λ > −λ1 that solves
φ1 (λ) p(λ) − 0.
(4.42)
85
86
CHAPTER 4.
TRUST-REGION METHODS
The disadvantage of this approach can be seen by considering the form of p(λ) when λ
is greater than, but close to, −λ1 . For such λ, we can approximate φ1 by a rational function,
as follows:
φ1 (λ) ≈
C1
+ C2 ,
λ + λ1
where C1 > 0 and C2 are constants. Clearly this approximation (and hence φ1 ) is highly
nonlinear, so the root-finding Newton’s method will be unreliable or slow. Better results will
be obtained if we reformulate the problem (4.42) so that it is nearly linear near the optimal
λ. By defining
φ2 (λ) 1
1
−
,
p(λ)
it can be shown using (4.39) that for λ slightly greater than −λ1 , we have
φ2 (λ) ≈
λ + λ1
1
−
C3
for some C3 > 0. Hence, φ2 is nearly linear near −λ1 (see Figure 4.6), and the root-finding
||p|| --1
1
3
Figure 4.6
2
1/ p(λ) as a function of λ.
1
*
4.3.
ITERATIVE SOLUTION
OF THE
SUBPROBLEM
Newton’s method will perform well, provided that it maintains λ > −λ1 . The root-finding
Newton’s method applied to φ2 generates a sequence of iterates λ() by setting
λ
(+1)
λ
()
φ2 λ()
− () .
φ2 λ
(4.43)
After some elementary manipulation, this updating formula can be implemented in the
following practical way.
Algorithm 4.3 (Trust Region Subproblem).
Given λ(0) , > 0:
for 0, 1, 2, . . .
Factor B + λ() I R T R;
Solve R T Rp −g, R T q p ;
Set
λ(+1) λ() +
p q 2 p − ;
(4.44)
end (for).
Safeguards must be added to this algorithm to make it practical; for instance, when
λ() < −λ1 , the Cholesky factorization B + λ() I R T R will not exist. A slightly enhanced
version of this algorithm does, however, converge to a solution of (4.37) in most cases.
The main work in each iteration of this method is, of course, the Cholesky factorization
of B + λ() I . Practical versions of this algorithm do not iterate until convergence to the
optimal λ is obtained with high accuracy, but are content with an approximate solution that
can be obtained in two or three iterations.
THE HARD CASE
Recall that in the discussion above, we assumed that q1T g 0. In fact, the approach
described above can be applied even when the most negative eigenvalue is a multiple
eigenvalue (that is, 0 > λ1 λ2 · · ·), provided that Q 1T g 0, where Q 1 is the matrix
whose columns span the subspace corresponding to the eigenvalue λ1 . When this condition
does not hold, the situation becomes a little complicated, because the limit (4.41) does not
hold for λ j λ1 and so there may not be a value λ ∈ (−λ1 , ∞) such that p(λ) (see
Figure 4.7). Moré and Sorensen [214] refer to this case as the hard case. At first glance, it is
not clear how p and λ can be chosen to satisfy (4.8) in the hard case. Clearly, our root-finding
technique will not work, since there is no solution for λ in the open interval (−λ1 , ∞). But
Theorem 4.1 assures us that the right value of λ lies in the interval [−λ1 , ∞), so there is only
87
88
CHAPTER 4.
TRUST-REGION METHODS
||p||
∆
−λ3
Figure 4.7
−λ2
−λ1
The hard case: p(λ) < for all λ ∈ (−λ1 , ∞).
one possibility: λ −λ1 . To find p, it is not enough to delete the terms for which λ j λ1
from the formula (4.38) and set
p
q Tj g
j:λ j λ1
λj + λ
qj.
Instead, we note that (B − λ1 I ) is singular, so there is a vector z such that z 1 and
(B − λ1 I )z 0. In fact, z is an eigenvector of B corresponding to the eigenvalue λ1 , so by
orthogonality of Q we have q Tj z 0 for λ j λ1 . It follows from this property that if we set
p
j:λ j λ1
q Tj g
λj + λ
qj + τz
(4.45)
for any scalar τ , we have
p2 j:λ j λ1
2
q Tj g
(λ j + λ)2
+ τ 2,
so it is always possible to choose τ to ensure that p . It is easy to check that the
conditions (4.8) holds for this choice of p and λ −λ1 .
4.3.
ITERATIVE SOLUTION
OF THE
SUBPROBLEM
PROOF OF THEOREM 4.1
We now give a formal proof of Theorem 4.1, the result that characterizes the exact
solution of (4.5). The proof relies on the following technical lemma, which deals with the
unconstrained minimizers of quadratics and is particularly interesting in the case where the
Hessian is positive semidefinite.
Lemma 4.7.
Let m be the quadratic function defined by
m( p) g T p + 12 p T Bp,
(4.46)
where B is any symmetric matrix. Then the following statements are true.
(i) m attains a minimum if and only if B is positive semidefinite and g is in the range of B.
If B is positive semidefinite, then every p satisfying Bp −g is a global minimizer of m.
(ii) m has a unique minimizer if and only if B is positive definite.
PROOF. We prove each of the three claims in turn.
(i) We start by proving the “if ” part. Since g is in the range of B, there is a p with Bp −g.
For all w ∈ R n , we have
m( p + w) g T ( p + w) + 12 ( p + w)T B( p + w)
(g T p + 12 p T Bp) + g T w + (Bp)T w + 12 w T Bw
m( p) + 12 w T Bw
≥ m( p),
(4.47)
since B is positive semidefinite. Hence, p is a minimizer of m.
For the “only if ” part, let p be a minimizer of m. Since ∇m( p) Bp + g 0, we
have that g is in the range of B. Also, we have ∇ 2 m( p) B positive semidefinite, giving
the result.
(ii) For the “if ” part, the same argument as in (i) suffices with the additional point that
w T Bw > 0 whenever w 0. For the “only if ” part, we proceed as in (i) to deduce that B is
positive semidefinite. If B is not positive definite, there is a vector w 0 such that Bw 0.
Hence, from (4.47), we have m( p + w) m( p), so the minimizer is not unique, giving a
contradiction.
To illustrate case (i), suppose that
⎡
1
⎢
B⎣ 0
0
0
0
⎤
0
⎥
0 ⎦,
0
2
89
90
CHAPTER 4.
TRUST-REGION METHODS
which has eigenvalues 0, 1, 2 and is therefore singular. If g is any vector whose second
component is zero, then g will be in the range of B, and the quadratic will attain a minimum.
But if the second element in g is nonzero, we can decrease m(·) indefinitely by moving along
the direction (0, −g2 , 0)T .
We are now in a position to take account of the trust-region bound p ≤ and
hence prove Theorem 4.1.
PROOF. (Theorem 4.1)
Assume first that there is λ ≥ 0 such that the conditions (4.8) are satisfied.
Lemma 4.7(i) implies that p ∗ is a global minimum of the quadratic function
m̂( p) g T p + 12 p T (B + λI ) p m( p) +
λ T
p p.
2
(4.48)
Since m̂( p) ≥ m̂( p ∗ ), we have
m( p) ≥ m( p ∗ ) +
λ
(( p ∗ )T p ∗ − p T p).
2
(4.49)
Because λ( − p ∗ ) 0 and therefore λ(2 − ( p ∗ )T p ∗ ) 0, we have
m( p) ≥ m( p ∗ ) +
λ 2
( − p T p).
2
Hence, from λ ≥ 0, we have m( p) ≥ m( p ∗ ) for all p with p ≤ . Therefore, p ∗ is a
global minimizer of (4.7).
For the converse, we assume that p ∗ is a global solution of (4.7) and show that there
is a λ ≥ 0 that satisfies (4.8).
In the case p ∗ < , p ∗ is an unconstrained minimizer of m, and so
∇m( p ∗ ) Bp ∗ + g 0,
∇ 2 m( p ∗ ) B positive semidefinite,
and so the properties (4.8) hold for λ 0.
Assume for the remainder of the proof that p ∗ . Then (4.8b) is immediately
satisfied, and p ∗ also solves the constrained problem
min m( p)
subject to p .
By applying optimality conditions for constrained optimization to this problem (see
(12.34)), we find that there is a λ such that the Lagrangian function defined by
L( p, λ) m( p) +
λ T
( p p − 2 )
2
4.3.
ITERATIVE SOLUTION
OF THE
SUBPROBLEM
has a stationary point at p ∗ . By setting ∇ p L( p ∗ , λ) to zero, we obtain
Bp ∗ + g + λp ∗ 0 ⇒ (B + λI ) p ∗ −g,
(4.50)
so that (4.8a) holds. Since m( p) ≥ m( p ∗ ) for any p with p T p ( p ∗ )T p ∗ 2 , we have
for such vectors p that
m( p) ≥ m( p ∗ ) +
λ ∗ T ∗
( p ) p − pT p .
2
If we substitute the expression for g from (4.50) into this expression, we obtain after some
rearrangement that
1
(p
2
− p ∗ )T (B + λI )( p − p ∗ ) ≥ 0.
(4.51)
Since the set of directions
p − p∗
w : w±
, for some p with p p − p∗ is dense on the unit sphere, (4.51) suffices to prove (4.8c).
It remains to show that λ ≥ 0. Because (4.8a) and (4.8c) are satisfied by p ∗ , we have
from Lemma 4.7(i) that p ∗ minimizes m̂, so (4.49) holds. Suppose that there are only negative
values of λ that satisfy (4.8a) and (4.8c). Then we have from (4.49) that m( p) ≥ m( p ∗ )
whenever p ≥ p ∗ . Since we already know that p ∗ minimizes m for p ≤ ,
it follows that m is in fact a global, unconstrained minimizer of m. From Lemma 4.7(i) it
follows that Bp −g and B is positive semidefinite. Therefore conditions (4.8a) and (4.8c)
are satisfied by λ 0, which contradicts our assumption that only negative values of λ can
satisfy the conditions. We conclude that λ ≥ 0, completing the proof.
CONVERGENCE OF ALGORITHMS BASED ON NEARLY EXACT SOLUTIONS
As we noted in the discussion of Algorithm 4.3, the loop to determine the optimal
values of λ and p for the subproblem (4.5) does not iterate until high accuracy is achieved.
Instead, it is terminated after two or three iterations with a fairly loose approximation to
the true solution. The inexactness in this approximate solution is measured in a different
way from the dogleg and subspace minimization algorithms. We can add safeguards to the
root-finding Newton method to ensure that the key assumptions of Theorems 4.5 and 4.6
are satisfied by the approximate solution. Specifically, we require that
m(0) − m( p) ≥ c1 (m(0) − m( p ∗ )),
p ≤ γ (4.52a)
(4.52b)
91
92
CHAPTER 4.
TRUST-REGION METHODS
(where p ∗ is the exact solution of (4.3)), for some constants c1 ∈ (0, 1] and γ > 0. The
condition (4.52a) ensures that the approximate solution achieves a significant fraction of the
maximum decrease possible in the model function m. (It is not necessary to know p ∗ ; there
are practical termination criteria that imply (4.52a).) One major difference between (4.52)
and the earlier criterion (4.20) is that (4.52) makes better use of the second-order part of
m(·), that is, the p T Bp term. This difference is illustrated by the case in which g 0 while
B has negative eigenvalues, indicating that the current iterate xk is a saddle point. Here,
the right-hand-side of (4.20) is zero (indeed, the algorithms we described earlier would
terminate at such a point). The right-hand-side of (4.52) is positive, indicating that decrease
in the model function is still possible, so it forces the algorithm to move away from xk .
The close attention that near-exact algorithms pay to the second-order term is warranted only if this term closely reflects the actual behavior of the function f —in fact,
the trust-region Newton method, for which B ∇ 2 f (x), is the only case that has been
treated in the literature. For purposes of global convergence analysis, the use of the exact
Hessian allows us to say more about the limit points of the algorithm than merely that they
are stationary points. The following result shows that second-order necessary conditions
(Theorem 2.3) are satisfied at the limit points.
Theorem 4.8.
Suppose that the assumptions of Theorem 4.6 are satisfied and in addition that f is twice
continuously differentiable in the level set S. Suppose that Bk ∇ 2 f (xk ) for all k, and that the
approximate solution pk of (4.3) at each iteration satisfies (4.52) for some fixed γ > 0. Then
limk→∞ gk 0.
If, in addition, the level set S of (4.24) is compact, then either the algorithm terminates
at a point xk at which the second-order necessary conditions (Theorem 2.3) for a local solution
hold, or else {xk } has a limit point x ∗ in S at which the second-order necessary conditions hold.
We omit the proof, which can be found in Moré and Sorensen [214, Section 4].
4.4
LOCAL CONVERGENCE OF TRUST-REGION NEWTON
METHODS
Since global convergence of trust-region methods that use exact Hessians ∇ 2 f (x) is established above, we turn our attention now to local convergence issues. The key to attaining
the fast rate of convergence usually associated with Newton’s method is to show that the
trust-region bound eventually does not interfere as we approach a solution. Specifically, we
hope that near the solution, the (approximate) solution of the trust-region subproblem is
well inside the trust region and becomes closer and closer to the true Newton step. Steps
that satisfy the latter property are said to be asymptotically similar to Newton steps.
We first prove a general result that applies to any algorithm of the form of Algorithm 4.1 (see Chapter 4) that generates steps that are asymptotically similar to Newton
4.4.
LOCAL CONVERGENCE
OF
TRUST-REGION NEWTON METHODS
steps whenever the Newton steps easily satisfy the trust-region bound. It shows that the
trust-region constraint eventually becomes inactive in algorithms with this property and
that superlinear convergence can be attained. The result assumes that the exact Hessian
Bk ∇ 2 f (xk ) is used in (4.3) when xk is close to a solution x ∗ that satisfies second-order
sufficient conditions (see Theorem 2.4). Moreover, it assumes that the algorithm uses an
approximate solution pk of (4.3) that achieves a similar decrease in the model function m k
as the Cauchy point.
Theorem 4.9.
Let f be twice Lipschitz continuously differentiable in a neighborhhod of a point x ∗ at
which second-order sufficient conditions (Theorem 2.4) are satisfied. Suppose the sequence {xk }
converges to x ∗ and that for all k sufficiently large, the trust-region algorithm based on (4.3)
with Bk ∇ 2 f (xk ) chooses steps pk that satisfy the Cauchy-point-based model reduction
criterion (4.20) and are asymptotically similar to Newton steps pkN whenever pkN ≤ 12 k ,
that is,
pk − pkN o( pkN ).
(4.53)
Then the trust-region bound k becomes inactive for all k sufficiently large and the sequence
{xk } converges superlinearly to x ∗ .
PROOF. We show that pkN ≤ 12 k and pk ≤ k , for all sufficiently large k, so the
near-optimal step pk in (4.53) will eventually always be taken.
We first seek a lower bound on the predicted reduction m k (0) − m k ( pk ) for all
sufficiently large k. We assume that k is large enough that the o( pkN ) term in (4.53) is less
than pkN . When pkN ≤ 12 k , we then have that pk ≤ pkN + o( pkN ) ≤ 2 pkN , while
if pkN > 12 k , we have pk ≤ k < 2 pkN . In both cases, then, we have
−1 pk ≤ 2 pkN ≤ 2 ∇ 2 f (xk ) gk ,
and so gk ≥ 12 pk / ∇ 2 f (xk )−1 .
We have from the relation (4.20) that
m k (0) − m k ( pk )
≥ c1 gk min
pk gk k , 2
∇ f (xk )
min
≥ c1 2 ∇ 2 f (xk )−1 pk pk , 2
2 ∇ f (xk ) ∇ 2 f (xk )−1 pk 2
c1 .
2 4 ∇ 2 f (xk )−1 ∇ 2 f (xk )
Because xk → x ∗ , we use continuity of ∇ 2 f (x) and positive definiteness of ∇ 2 f (x ∗ ), to
93
94
CHAPTER 4.
TRUST-REGION METHODS
deduce that the following bound holds for all k sufficiently large:
c1
c1
def
≥ c3 ,
2 2 −1
−1
2
2
2
∗
2
∗
4 ∇ f (xk ) ∇ f (xk )
8 ∇ f (x ) ∇ f (x )
where c3 > 0. Hence, we hae
m k (0) − m k ( pk ) ≥ c3 pk 2
(4.54)
for all sufficiently large k. By Lipschitz continuity of ∇ 2 f (x) near x ∗ , and using Taylor’s
theorem (Theorem 2.1), we have
|( f (xk ) − f (xk + pk )) − (m k (0) − m k ( pk ))|
1
1 T 2
T 2
1
pk ∇ f (xk + t pk ) pk dt 2 pk ∇ f (xk ) pk − 2
0
L
≤ pk 3 ,
4
where L > 0 is the Lipschitz constant for ∇ 2 f (·). Hence, by definition (4.4) of ρk , we have
for sufficiently large k that
|ρk − 1| ≤
pk 3 (L/4)
L
L
pk ≤
k .
c3 pk 2
4c3
4c3
(4.55)
Now, the trust-region radius can be reduced only if ρk < 14 (or some other fixed number less
than 1), so it is clear from (4.55) that the sequence {k } is bounded away from zero. Since
xk → x ∗ , we have pkN → 0 and therefore pk → 0 from (4.53). Hence, the trust-region
bound is inactive for all k sufficiently large, and the bound pkN ≤ 12 k is eventually always
satisfied.
To prove superlinear convergence, we use the quadratic convergence of Newton’s
method, proved in Theorem 3.5. In particular, we have from (3.33) that
xk + pkN − x ∗ o xk − x ∗ 2 ,
which implies that pkN O(xk − x ∗ ). Therefore, using (4.53), we have
xk + pk − x ∗ ≤ xk + pkN − x ∗ + pkN − pk o xk − x ∗ 2 + o( pkN ) o xk − x ∗ ,
thus proving superlinear convergence.
It is immediate from Theorem 3.5 that if pk pkN for all k sufficiently large, we have
quadratic convergence of {xk } to x ∗ .
4.5.
OTHER ENHANCEMENTS
Reasonable implementations of the dogleg, subspace minimization, and nearly-exact
algorithm of Section 4.3 with Bk ∇ 2 f (xk ) eventually use the steps pk pkN under the
conditions of Theorem 4.9, and therefore converge quadratically. In the case of the dogleg and
two-dimensional subspace minimization methods, the exact step pkN is one of the candidates
for pk —it lies inside the trust region, along the dogleg path, and inside the two-dimensional
subspace. Since under the assumptions of Theorem 4.9, pkN is the unconstrained minimizer
of m k for k sufficiently large, it is certainly the minimizer in the more restricted domains,
so we have pk pkN . For the approach of Section 4.3, if we follow the reasonable strategy
of checking whether pkN is a solution of (4.3) prior to embarking on Algorithm 4.3, then
eventually we will also have pk pkN also.
4.5
OTHER ENHANCEMENTS
SCALING
As we noted in Chapter 2, optimization problems are often posed with poor scaling—
the objective function f is highly sensitive to small changes in certain components of
the vector x and relatively insensitive to changes in other components. Topologically, a
symptom of poor scaling is that the minimizer x ∗ lies in a narrow valley, so that the contours
of the objective f (·) near x ∗ tend towards highly eccentric ellipses. Algorithms that fail to
compensate for poor scaling can perform badly; see Figure 2.7 for an illustration of the poor
performance of the steepest descent approach.
Recalling our definition of a trust region—a region around the current iterate within
which the model m k (·) is an adequate representation of the true objective f (·)—it is easy
to see that a spherical trust region may not be appropriate when f is poorly scaled. Even if
the model Hessian Bk is exact, the rapid changes in f along certain directions probably will
cause m k to be a poor approximation to f along these directions. On the other hand, m k
may be a more reliable approximation to f along directions in which f is changing more
slowly. Since the shape of our trust region should be such that our confidence in the model
is more or less the same at all points on the boundary of the region, we are led naturally
to consider elliptical trust regions in which the axes are short in the sensitive directions and
longer in the less sensitive directions.
Elliptical trust regions can be defined by
Dp ≤ ,
(4.56)
where D is a diagonal matrix with positive diagonal elements, yielding the following scaled
trust-region subproblem:
def
minn m k ( p) f k + gkT p + 12 p T Bk p
p∈IR
s.t. Dp ≤ k .
(4.57)
95
96
CHAPTER 4.
TRUST-REGION METHODS
When f (x) is highly sensitive to the value of the ith component xi , we set the corresponding
diagonal element dii of D to be large, while dii is smaller for less-sensitive components.
Information to construct the scaling matrix D may be derived from the second
derivatives ∂ 2 f /∂ xi2 . We can allow D to change from iteration to iteration; most of the
theory of this chapter will still apply with minor modifications provided that each dii stays
within some predetermined range [dlo , dhi ], where 0 < dlo ≤ dhi < ∞. Of course, we do
not need D to be a precise reflection of the scaling of the problem, so it is not necessary to
devise elaborate heuristics or to perform extensive computations to get it just right.
The following procedure shows how the Cauchy point calculation (Algorithm 4.2)
changes when we use a scaled trust region,
Algorithm 4.4 (Generalized Cauchy Point Calculation).
Find the vector pkS that solves
pkS arg minn f k + gkT p
s.t. Dp ≤ k ;
p∈IR
(4.58)
Calculate the scalar τk > 0 that minimizes m k (τ pkS ) subject to satisfying the trust-region
bound, that is,
τk arg min m k (τ pkS )
τ >0
s.t. τ D pkS ≤ k ;
(4.59)
pkC τk pkS .
For this scaled version, we find that
pkS −
k
D −2 gk ,
D −1 gk (4.60)
and that the step length τk is obtained from the following modification of (4.12):
τk ⎧
⎪
⎨ 1
⎪
⎩ min
D −1 gk 3
k gkT D −2 Bk D −2 gk
,1
if gkT D −2 Bk D −2 gk ≤ 0
otherwise.
(4.61)
(The details are left as an exercise.)
A simpler alternative for adjusting the definition of the Cauchy point and the various
algorithms of this chapter to allow for the elliptical trust region is simply to rescale the
variables p in the subproblem (4.57) so that the trust region is spherical in the scaled
variables. By defining
def
p̃ Dp,
4.5.
OTHER ENHANCEMENTS
and by substituting into (4.57), we obtain
minn m̃ k ( p̃) f k + gkT D −1 p̃ + 12 p̃ T D −1 Bk D −1 p̃
def
p̃∈IR
s.t. p̃ ≤ k .
The theory and algorithms can now be derived in the usual way by substituting p̃ for p,
D −1 gk for gk , D −1 Bk D −1 for Bk , and so on.
TRUST REGIONS IN OTHER NORMS
Trust regions may also be defined in terms of norms other than the Euclidean norm.
For instance, we may have
p1 ≤ k
or
p∞ ≤ k ,
or
Dp∞ ≤ k ,
or their scaled counterparts
Dp1 ≤ k
where D is a positive diagonal matrix as before. Norms such as these offer no obvious advantages for small-medium unconstrained problems, but they may be useful for constrained
problems. For instance, for the bound-constrained problem
subject to x ≥ 0,
minn f (x),
x∈IR
the trust-region subproblem may take the form
minn m k ( p) f k + gkT p + 12 p T Bk p
p∈IR
s.t. xk + p ≥ 0, p ≤ k .
(4.62)
When the trust region is defined by a Euclidean norm, the feasible region for (4.62) consists of
the intersection of a sphere and the nonnegative orthant—an awkward object, geometrically
speaking. When the ∞-norm is used, however, the feasible region is simply the rectangular
box defined by
xk + p ≥ 0,
p ≥ −k e,
p ≤ k e,
where e (1, 1, . . . , 1)T , so the solution of the subproblem is easily calculated by using
techniques for bound-constrained quadratic programming.
For large problems, in which factorization or formation the model Hessian Bk is not
computationally desirable, the use of a trust region defined by · ∞ will also give rise to a
bound-constrained subproblem, which may be more convenient to solve than the standard
subproblem (4.3). To our knowledge, there has not been much research on the relative
performance of methods that use trust regions of different shapes on large problems.
97
98
CHAPTER 4.
TRUST-REGION METHODS
NOTES AND REFERENCES
One of the earliest works on trust-region methods is Winfield [307]. The influential
paper of Powell [244] proves a result like Theorem 4.5 for the case of η 0, where the algorithm takes a step whenever it decreases the function value. Powell uses a weaker assumption
than ours on the matrices B, but his analysis is more complicated. Moré [211] summarizes
developments in algorithms and software before 1982, paying particular attention to the
importance of using a scaled trust-region norm.
Byrd, Schnabel, and Schultz [279], [54] provide a general theory for inexact trustregion methods; they introduce the idea of two-dimensional subspace minimization and
also focus on proper handling of the case of indefinite B to ensure stronger local convergence
results than Theorems 4.5 and 4.6. Dennis and Schnabel [93] survey trust-region methods as
part of their overview of unconstrained optimization, providing pointers to many important
developments in the literature.
The monograph of Conn, Gould, and Toint [74] is an exhaustive treatment of the state
of the art in trust-region methods for both unconstrained and constrained optimization. It
includes an comprehensive annotated bibliography of the literature in the area.
✐
EXERCISES
✐
4.1 Let f (x) 10(x2 − x12 )2 + (1 − x1 )2 . At x (0, −1) draw the contour lines of
the quadratic model (4.2) assuming that B is the Hessian of f . Draw the family of solutions
of (4.3) as the trust region radius varies from 0 to 2. Repeat this at x (0, 0.5).
✐
4.2 Write a program that implements the dogleg method. Choose Bk to be the exact
Hessian. Apply it to solve Rosenbrock’s function (2.22). Experiment with the update rule
for the trust region by changing the constants in Algorithm 4.1, or by designing your own
rules.
✐
4.3 Program the trust-region method based on Algorithm 7.2. Choose Bk to be the
exact Hessian, and use it to minimize the function
min f (x) n
2
(1 − x2i−1 )2 + 10(x2i − x2i−1
)2
i1
with n 10. Experiment with the starting point and the stopping test for the CG iteration.
Repeat the computation with n 50.
Your program should indicate, at every iteration, whether Algorithm 7.2 encountered
negative curvature, reached the trust-region boundary, or met the stopping test.
4.5.
OTHER ENHANCEMENTS
✐
4.4 Theorem 4.5 shows that the sequence {g} has an accumulation point at zero.
Show that if the iterates x stay in a bounded set B, then there is a limit point x∞ of the
sequence {xk } such that g(x∞ ) 0.
✐
4.5 Show that τk defined by (4.12) does indeed identify the minimizer of m k along
the direction −gk .
✐
4.6 The Cauchy–Schwarz inequality states that for any vectors u and v, we have
|u T v|2 ≤ (u T u)(v T v),
with equality only when u and v are parallel. When B is positive definite, use this inequality
to show that
def
γ g4
≤ 1,
(g T Bg)(g T B −1 g)
with equality only if g and Bg (and B −1 g) are parallel.
✐
4.7 When B is positive definite, the double-dogleg method constructs a path with three
line segments from the origin to the full step. The four points that define the path are
• the origin;
• the unconstrained Cauchy step p C −(g T g)/(g T Bg)g;
• a fraction of the full step γ̄ p B −γ̄ B −1 g, for some γ̄ ∈ (γ , 1], where γ is defined in
the previous question; and
• the full step p B −B −1 g.
Show that p increases monotonically along this path.
(Note: The double-dogleg method, as discussed in Dennis and Schnabel [92, Section
6.4.2], was for some time thought to be superior to the standard dogleg method, but later
testing has not shown much difference in performance.)
✐
4.8 Show that (4.43) and (4.44) are equivalent. Hints: Note that
d
dλ
−1/2
−3/2 d
d 1
p(λ)2
p(λ)2 ,
− p(λ)2
dλ
2
dλ
n
(q Tj g)2
d
p(λ)2 −2
dλ
(λ j + λ)3
j1
1
p(λ)
99
100
CHAPTER 4.
TRUST-REGION METHODS
(from (4.39)), and
q2 R −T p2 p T (B + λI )−1 p n
(q Tj g)2
j1
(λ j + λ)3
.
✐
4.9 Derive the solution of the two-dimensional subspace minimization problem in
the case where B is positive definite.
✐
4.10 Show that if B is any symmetric matrix, then there exists λ ≥ 0 such that B + λI
is positive definite.
✐
4.11 Verify that the definitions (4.60) for pkS and (4.61) for τk are valid for the Cauchy
point in the case of an elliptical trust region. (Hint: Using the theory of Chapter 12, we can
show that the solution of (4.58) satisfies gk + α D 2 pkS 0 for some scalar α ≥ 0.)
✐
4.12 The following example shows that the reduction in the model function m
achieved by the two-dimensional minimization strategy can be much smaller than that
achieved by the exact solution of (4.5).
In (4.5), set
T
1
g − , −1, −
2
,
where is a small positive number. Set
B diag
1
3
, 1,
3
,
0.5.
T
Show that the solution of (4.5) has components O( ), 12 + O( ), O( ) and that the
reduction in the model m is 38 + O( ). For the two-dimensional minimization strategy,
show that the solution is a multiple of B −1 g and that the reduction in m is O( ).
This is page 101
Printer: Opaque this
CHAPTER
5
Conjugate
Gradient Methods
Our interest in conjugate gradient methods is twofold. First, they are among the most useful
techniques for solving large linear systems of equations. Second, they can be adapted to solve
nonlinear optimization problems. The remarkable properties of both linear and nonlinear
conjugate gradient methods will be described in this chapter.
The linear conjugate gradient method was proposed by Hestenes and Stiefel in the
1950s as an iterative method for solving linear systems with positive definite coefficient
matrices. It is an alternative to Gaussian elimination that is well suited for solving large
problems. The performance of the linear conjugate gradient method is determined by the
102
CHAPTER 5.
CONJUGATE GRADIENT METHODS
distribution of the eigenvalues of the coefficient matrix. By transforming, or preconditioning,
the linear system, we can make this distribution more favorable and improve the convergence
of the method significantly. Preconditioning plays a crucial role in the design of practical
conjugate gradient strategies. Our treatment of the linear conjugate gradient method will
highlight those properties of the method that are important in optimization.
The first nonlinear conjugate gradient method was introduced by Fletcher and Reeves
in the 1960s. It is one of the earliest known techniques for solving large-scale nonlinear
optimization problems. Over the years, many variants of this original scheme have been
proposed, and some are widely used in practice. The key features of these algorithms are
that they require no matrix storage and are faster than the steepest descent method.
5.1
THE LINEAR CONJUGATE GRADIENT METHOD
In this section we derive the linear conjugate gradient method and discuss its essential
convergence properties. For simplicity, we drop the qualifier “linear” throughout.
The conjugate gradient method is an iterative method for solving a linear system of
equations
Ax b,
(5.1)
where A is an n × n symmetric positive definite matrix. The problem (5.1) can be stated
equivalently as the following minimization problem:
def
min φ(x) 12 x T Ax − b T x,
(5.2)
that is, both (5.1) and (5.2) have the same unique solution. This equivalence will allow us
to interpret the conjugate gradient method either as an algorithm for solving linear systems
or as a technique for minimizing convex quadratic functions. For future reference, we note
that the gradient of φ equals the residual of the linear system, that is,
def
∇φ(x) Ax − b r (x),
(5.3)
so in particular at x xk we have
rk Axk − b.
(5.4)
CONJUGATE DIRECTION METHODS
One of the remarkable properties of the conjugate gradient method is its ability to
generate, in a very economical fashion, a set of vectors with a property known as conjugacy. A
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
set of nonzero vectors { p0 , p1 , . . . , pl } is said to be conjugate with respect to the symmetric
positive definite matrix A if
piT Ap j 0,
for all i j.
(5.5)
It is easy to show that any set of vectors satisfying this property is also linearly independent.
(For a geometrical illustration of conjugate directions see Section 9.4.)
The importance of conjugacy lies in the fact that we can minimize φ(·) in n steps
by successively minimizing it along the individual directions in a conjugate set. To verify
this claim, we consider the following conjugate direction method. (The distinction between
the conjugate gradient method and the conjugate direction method will become clear as we
proceed.) Given a starting point x0 ∈ IRn and a set of conjugate directions { p0 , p1 , . . . , pn−1 },
let us generate the sequence {xk } by setting
xk+1 xk + αk pk ,
(5.6)
where αk is the one-dimensional minimizer of the quadratic function φ(·) along xk + αpk ,
given explicitly by
αk −
rkT pk
;
pkT Apk
(5.7)
see (3.55). We have the following result.
Theorem 5.1.
For any x0 ∈ IRn the sequence {xk } generated by the conjugate direction algorithm (5.6),
(5.7) converges to the solution x ∗ of the linear system (5.1) in at most n steps.
PROOF. Since the directions { pi } are linearly independent, they must span the whole space
IRn . Hence, we can write the difference between x0 and the solution x ∗ in the following way:
x ∗ − x0 σ0 p0 + σ1 p1 + · · · + σn−1 pn−1 ,
for some choice of scalars σk . By premultiplying this expression by pkT A and using the
conjugacy property (5.5), we obtain
σk pkT A(x ∗ − x0 )
.
pkT Apk
(5.8)
We now establish the result by showing that these coefficients σk coincide with the step
lengths αk generated by the formula (5.7).
103
104
CHAPTER 5.
CONJUGATE GRADIENT METHODS
If xk is generated by algorithm (5.6), (5.7), then we have
xk x0 + α0 p0 + α1 p1 + · · · + αk−1 pk−1 .
By premultiplying this expression by pkT A and using the conjugacy property, we have that
pkT A(xk − x0 ) 0,
and therefore
pkT A(x ∗ − x0 ) pkT A(x ∗ − xk ) pkT (b − Axk ) − pkT rk .
By comparing this relation with (5.7) and (5.8), we find that σk αk , giving the result.
There is a simple interpretation of the properties of conjugate directions. If the matrix
A in (5.2) is diagonal, the contours of the function φ(·) are ellipses whose axes are aligned
with the coordinate directions, as illustrated in Figure 5.1. We can find the minimizer of this
function by performing one-dimensional minimizations along the coordinate directions
e2
.
x*
.
x0
.
x1
e1
Figure 5.1 Successive minimizations along the coordinate directions find the
minimizer of a quadratic with a diagonal Hessian in n iterations.
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
e2
x*
x3
x2
x0
x1
e1
Figure 5.2 Successive minimization along coordinate axes does not find the solution
in n iterations, for a general convex quadratic.
e1 , e2 , . . . , en in turn. When A is not diagonal, its contours are still elliptical, but they
are usually no longer aligned with the coordinate directions. The strategy of successive
minimization along these directions in turn no longer leads to the solution in n iterations (or
even in a finite number of iterations). This phenomenon is illustrated in the two-dimensional
example of Figure 5.2 We can, however, recover the nice behavior of Figure 5.1 if we transform
the problem to make A diagonal and then minimize along the coordinate directions. Suppose
we transform the problem by defining new variables x̂ as
x̂ S −1 x,
(5.9)
where S is the n × n matrix defined by
S [ p0 p1 · · · pn−1 ],
where { p0 , p2 , . . . , pn−1 } is the set of conjugate directions with respect to A. The quadratic
φ defined by (5.2) now becomes
def
φ̂(x̂) φ(S x̂) 12 x̂ T (S T AS)x̂ − (S T b)T x̂.
By the conjugacy property (5.5), the matrix S T AS is diagonal, so we can find the minimizing
value of φ̂ by performing n one-dimensional minimizations along the coordinate directions
105
106
CHAPTER 5.
CONJUGATE GRADIENT METHODS
of x̂. Because of the relation (5.9), however, the ith coordinate direction in x̂-space corresponds to the direction pi in x-space. Hence, the coordinate search strategy applied to φ̂ is
equivalent to the conjugate direction algorithm (5.6), (5.7). We conclude, as in Theorem 5.1,
that the conjugate direction algorithm terminates in at most n steps.
Returning to Figure 5.1, we note another interesting property: When the Hessian matrix is diagonal, each coordinate minimization correctly determines one of the components
of the solution x ∗ . In other words, after k one-dimensional minimizations, the quadratic
has been minimized on the subspace spanned by e1 , e2 , . . . , ek . The following theorem
proves this important result for the general case in which the Hessian of the quadratic is not
necessarily diagonal. (Here and later, we use the notation span{ p0 , p1 , . . . , pk } to denote
the set of all linear combinations of the vectors p0 , p1 , . . . , pk .) In proving the result we will
make use of the following expression, which is easily verified from the relations (5.4) and
(5.6):
rk+1 rk + αk Apk .
(5.10)
Theorem 5.2 (Expanding Subspace Minimization).
Let x0 ∈ IRn be any starting point and suppose that the sequence {xk } is generated by the
conjugate direction algorithm (5.6), (5.7). Then
rkT pi 0,
for i 0, 1, . . . , k − 1,
(5.11)
and xk is the minimizer of φ(x) 12 x T Ax − b T x over the set
{x | x x0 + span{ p0 , p1 , . . . , pk−1 }}.
(5.12)
PROOF. We begin by showing that a point x̃ minimizes φ over the set (5.12) if and only
if r (x̃)T pi 0, for each i 0, 1, . . . , k − 1. Let us define h(σ ) φ(x0 + σ0 p0 + · · · +
σk−1 pk−1 ), where σ (σ0 , σ1 , . . . , σk−1 )T . Since h(σ ) is a strictly convex quadratic, it has
a unique minimizer σ ∗ that satisfies
∂h(σ ∗ )
0,
∂σi
i 0, 1, . . . , k − 1.
By the chain rule, this equation implies that
∗
∇φ(x0 + σ0∗ p0 + · · · + σk−1
pk−1 )T pi 0,
i 0, 1, . . . , k − 1.
By recalling the definition (5.3), we have for the minimizer x̃ x0 + σ0∗ p0 + σ1∗ p2 + · · · +
∗
pk−1 on the set (5.12) that r (x̃)T pi 0, as claimed.
σk−1
We now use induction to show that xk satisfies (5.11). For the case k 1, we have
from the fact that x1 x0 + α0 p0 minimizes φ along p0 that r1T p0 0. Let us now make
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
T
the induction hypothesis, namely, that rk−1
pi 0 for i 0, 1, . . . , k − 2. By (5.10), we
have
rk rk−1 + αk−1 Apk−1 ,
so that
T
T
T
rk pk−1
rk−1 + αk−1 pk−1
Apk−1 0,
pk−1
by the definition (5.7) of αk−1 . Meanwhile, for the other vectors pi , i 0, 1, . . . , k − 2, we
have
piT rk piT rk−1 + αk−1 piT Apk−1 0,
where piT rk−1 0 because of the induction hypothesis and piT Apk−1 0 because of
conjugacy of the vectors pi . We have shown that rkT pi 0, for i 0, 1, . . . , k − 1, so the
proof is complete.
The fact that the current residual rk is orthogonal to all previous search directions, as
expressed in (5.11), is a property that will be used extensively in this chapter.
The discussion so far has been general, in that it applies to a conjugate direction
method (5.6), (5.7) based on any choice of the conjugate direction set { p0 , p1 , . . . , pn−1 }.
There are many ways to choose the set of conjugate directions. For instance, the eigenvectors v1 , v2 , . . . , vn of A are mutually orthogonal as well as conjugate with respect to
A, so these could be used as the vectors { p0 , p1 , . . . , pn−1 }. For large-scale applications,
however, computation of the complete set of eigenvectors requires an excessive amount of
computation. An alternative approach is to modify the Gram–Schmidt orthogonalization
process to produce a set of conjugate directions rather than a set of orthogonal directions.
(This modification is easy to produce, since the properties of conjugacy and orthogonality
are closely related in spirit.) However, the Gram–Schmidt approach is also expensive, since
it requires us to store the entire direction set.
BASIC PROPERTIES OF THE CONJUGATE GRADIENT METHOD
The conjugate gradient method is a conjugate direction method with a very special
property: In generating its set of conjugate vectors, it can compute a new vector pk by
using only the previous vector pk−1 . It does not need to know all the previous elements
p0 , p1 , . . . , pk−2 of the conjugate set; pk is automatically conjugate to these vectors. This
remarkable property implies that the method requires little storage and computation.
In the conjugate gradient method, each direction pk is chosen to be a linear combination of the negative residual −rk (which, by (5.3), is the steepest descent direction for the
107
108
CHAPTER 5.
CONJUGATE GRADIENT METHODS
function φ) and the previous direction pk−1 . We write
pk −rk + βk pk−1 ,
(5.13)
where the scalar βk is to be determined by the requirement that pk−1 and pk must be
T
A and imposing the condition
conjugate with respect to A. By premultiplying (5.13) by pk−1
T
pk−1 Apk 0, we find that
βk rkT Apk−1
.
T
pk−1
Apk−1
We choose the first search direction p0 to be the steepest descent direction at the initial point
x0 . As in the general conjugate direction method, we perform successive one-dimensional
minimizations along each of the search directions. We have thus specified a complete
algorithm, which we express formally as follows:
Algorithm 5.1 (CG–Preliminary Version).
Given x0 ;
Set r0 ← Ax0 − b, p0 ← −r0 , k ← 0;
while rk 0
rkT pk
;
pkT Apk
← xk + αk pk ;
αk ← −
xk+1
rk+1 ← Axk+1 − b;
r T Apk
βk+1 ← k+1
;
pkT Apk
pk+1 ← −rk+1 + βk+1 pk ;
k ← k + 1;
(5.14a)
(5.14b)
(5.14c)
(5.14d)
(5.14e)
(5.14f)
end (while)
This version is useful for studying the essential properties of the conjugate gradient
method, but we present a more efficient version later. We show first that the directions
p0 , p1 , . . . , pn−1 are indeed conjugate, which by Theorem 5.1 implies termination in n
steps. The theorem below establishes this property and two other important properties.
First, the residuals ri are mutually orthogonal. Second, each search direction pk and residual
rk is contained in the Krylov subspace of degree k for r0 , defined as
def
K(r0 ; k) span{r0 , Ar0 , . . . , Ak r0 }.
(5.15)
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
Theorem 5.3.
Suppose that the kth iterate generated by the conjugate gradient method is not the solution
point x ∗ . The following four properties hold:
rkT ri 0,
for i 0, 1, . . . , k − 1,
(5.16)
span{r0 , r1 , . . . , rk } span{r0 , Ar0 , . . . , A r0 },
(5.17)
span{ p0 , p1 , . . . , pk } span{r0 , Ar0 , . . . , A r0 },
(5.18)
k
k
pkT
Api 0,
for i 0, 1, . . . , k − 1.
(5.19)
Therefore, the sequence {xk } converges to x ∗ in at most n steps.
PROOF. The proof is by induction. The expressions (5.17) and (5.18) hold trivially for k 0,
while (5.19) holds by construction for k 1. Assuming now that these three expressions are
true for some k (the induction hypothesis), we show that they continue to hold for k + 1.
To prove (5.17), we show first that the set on the left-hand side is contained in the set
on the right-hand side. Because of the induction hypothesis, we have from (5.17) and (5.18)
that
rk ∈ span{r0 , Ar0 , . . . , Ak r0 },
pk ∈ span{r0 , Ar0 , . . . , Ak r0 },
while by multiplying the second of these expressions by A, we obtain
Apk ∈ span{Ar0 , . . . , Ak+1 r0 }.
(5.20)
By applying (5.10), we find that
rk+1 ∈ span{r0 , Ar0 , . . . , Ak+1r0 }.
By combining this expression with the induction hypothesis for (5.17), we conclude that
span{r0 , r1 , . . . , rk , rk+1 } ⊂ span{r0 , Ar0 , . . . , Ak+1r0 }.
To prove that the reverse inclusion holds as well, we use the induction hypothesis on (5.18)
to deduce that
Ak+1r0 A(Ak r0 ) ∈ span{Ap0 , Ap1 , . . . , Apk }.
Since by (5.10) we have Api (ri+1 − ri )/αi for i 0, 1, . . . , k, it follows that
Ak+1r0 ∈ span{r0 , r1 , . . . , rk+1 }.
109
110
CHAPTER 5.
CONJUGATE GRADIENT METHODS
By combining this expression with the induction hypothesis for (5.17), we find that
span{r0 , Ar0 , . . . , Ak+1r0 } ⊂ span{r0 , r1 , . . . , rk , rk+1 }.
Therefore, the relation (5.17) continues to hold when k is replaced by k + 1, as claimed.
We show that (5.18) continues to hold when k is replaced by k + 1 by the following
argument:
span{ p0 , p1 , . . . , pk , pk+1 }
span{ p0 , p1 , . . . , pk , rk+1 }
by (5.14e)
span{r0 , Ar0 , . . . , A r0 , rk+1 }
by induction hypothesis for (5.18)
span{r0 , r1 , . . . , rk , rk+1 }
by (5.17)
span{r0 , Ar0 , . . . , A
by (5.17) for k + 1.
k
k+1
r0 }
Next, we prove the conjugacy condition (5.19) with k replaced by k +1. By multiplying
(5.14e) by Api , i 0, 1, . . . , k, we obtain
T
T
pk+1
Api −rk+1
Api + βk+1 pkT Api .
(5.21)
By the definition (5.14d) of βk , the right-hand-side of (5.21) vanishes when i k. For
i ≤ k − 1 we need to collect a number of observations. Note first that our induction
hypothesis for (5.19) implies that the directions p0 , p1 , . . . , pk are conjugate, so we can
apply Theorem 5.2 to deduce that
T
rk+1
pi 0,
for i 0, 1, . . . , k.
(5.22)
Second, by repeatedly applying (5.18), we find that for i 0, 1, . . . , k − 1, the following
inclusion holds:
Api ∈ A span{r0 , Ar0 , . . . , Ai r0 } span{Ar0 , A2r0 , . . . , Ai+1 r0 }
⊂ span{ p0 , p1 , . . . , pi+1 }.
(5.23)
By combining (5.22) and (5.23), we deduce that
T
rk+1
Api 0,
for i 0, 1, . . . , k − 1,
so the first term in the right-hand-side of (5.21) vanishes for i 0, 1, . . . , k − 1. Because of the induction hypothesis for (5.19), the second term vanishes as well, and we
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
T
conclude that pk+1
Api 0, i 0, 1, . . . , k. Hence, the induction argument holds for (5.19)
also.
It follows that the direction set generated by the conjugate gradient method is indeed
a conjugate direction set, so Theorem 5.1 tells us that the algorithm terminates in at most n
iterations.
Finally, we prove (5.16) by a noninductive argument. Because the direction set is
conjugate, we have from (5.11) that rkT pi 0 for all i 0, 1, . . . , k − 1 and any k 1, 2, . . . , n − 1. By rearranging (5.14e), we find that
pi −ri + βi pi−1 ,
so that ri ∈ span{ pi , pi−1 } for all i 1, . . . , k − 1. We conclude that rkT ri 0 for all
i 1, . . . , k − 1. To complete the proof, we note that rkT r0 −rkT p0 0, by definition of
p0 in Algorithm 5.1 and by (5.11).
The proof of this theorem relies on the fact that the first direction p0 is the steepest descent direction −r0 ; in fact, the result does not hold for other choices of p0 . Since
the gradients rk are mutually orthogonal, the term “conjugate gradient method” is actually a misnomer. It is the search directions, not the gradients, that are conjugate with
respect to A.
A PRACTICAL FORM OF THE CONJUGATE GRADIENT METHOD
We can derive a slightly more economical form of the conjugate gradient method by
using the results of Theorems 5.2 and 5.3. First, we can use (5.14e) and (5.11) to replace the
formula (5.14a) for αk by
αk rkT rk
.
pkT Apk
Second, we have from (5.10) that αk Apk rk+1 − rk , so by applying (5.14e) and (5.11)
once again we can simplify the formula for βk+1 to
βk+1 T
rk+1
rk+1
rkT rk
.
By using these formulae together with (5.10), we obtain the following standard form of the
conjugate gradient method.
111
112
CHAPTER 5.
CONJUGATE GRADIENT METHODS
Algorithm 5.2 (CG).
Given x0 ;
Set r0 ← Ax0 − b, p0 ← −r0 , k ← 0;
while rk 0
rkT rk
;
pkT Apk
← xk + αk pk ;
αk ←
xk+1
rk+1 ← rk + αk Apk ;
r T rk+1
βk+1 ← k+1T
;
rk rk
pk+1 ← −rk+1 + βk+1 pk ;
k ← k + 1;
(5.24a)
(5.24b)
(5.24c)
(5.24d)
(5.24e)
(5.24f)
end (while)
At any given point in Algorithm 5.2 we never need to know the vectors x, r , and
p for more than the last two iterations. Accordingly, implementations of this algorithm
overwrite old values of these vectors to save on storage. The major computational tasks to be
performed at each step are computation of the matrix–vector product Apk , calculation of
T
rk+1 , and calculation of three vector sums. The inner
the inner products pkT (Apk ) and rk+1
product and vector sum operations can be performed in a small multiple of n floating-point
operations, while the cost of the matrix–vector product is, of course, dependent on the
problem. The CG method is recommended only for large problems; otherwise, Gaussian
elimination or other factorization algorithms such as the singular value decomposition are
to be preferred, since they are less sensitive to rounding errors. For large problems, the CG
method has the advantage that it does not alter the coefficient matrix and (in contrast to
factorization techniques) does not produce fill in the arrays holding the matrix. Another key
property is that the CG method sometimes approaches the solution quickly, as we discuss
next.
RATE OF CONVERGENCE
We have seen that in exact arithmetic the conjugate gradient method will terminate at
the solution in at most n iterations. What is more remarkable is that when the distribution
of the eigenvalues of A has certain favorable features, the algorithm will identify the solution
in many fewer than n iterations. To explain this property, we begin by viewing the expanding
subspace minimization property proved in Theorem 5.2 in a slightly different way, using it
to show that Algorithm 5.2 is optimal in a certain important sense.
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
From (5.24b) and (5.18), we have that
xk+1 x0 + α0 p0 + · · · + αk pk
x0 + γ0r0 + γ1 Ar0 + · · · + γk Ak r0 ,
(5.25)
for some constants γi . We now define Pk∗ (·) to be a polynomial of degree k with coefficients
γ0 , γ1 , . . . , γk . Like any polynomial, Pk∗ can take either a scalar or a square matrix as its
argument. For the matrix argument A, we have
Pk∗ (A) γ0 I + γ1 A + · · · + γk Ak ,
which allows us to express (5.25) as follows:
xk+1 x0 + Pk∗ (A)r0 .
(5.26)
We now show that among all possible methods whose first k steps are restricted to the
Krylov subspace K(r0 ; k) given by (5.15), Algorithm 5.2 does the best job of minimizing the
distance to the solution after k steps, when this distance is measured by the weighted norm
measure · A defined by
z2A z T Az.
(5.27)
(Recall that this norm was used in the analysis of the steepest descent method of Chapter 3.)
Using this norm and the definition of φ (5.2), and the fact that x ∗ minimizes φ, it is easy to
show that
1
x
2
− x ∗ 2A 12 (x − x ∗ )T A(x − x ∗ ) φ(x) − φ(x ∗ ).
(5.28)
Theorem 5.2 states that xk+1 minimizes φ, and hence x − x ∗ 2A , over the set x0 +
span{ p0 , p1 , . . . , pk }, which by (5.18) is the same as x0 +span{r0 , Ar0 , . . . , Ak r0 }. It follows
from (5.26) that the polynomial Pk∗ solves the following problem in which the minimum is
taken over the space of all possible polynomials of degree k:
min x0 + Pk (A)r0 − x ∗ A .
Pk
(5.29)
We exploit this optimality property repeatedly in the remainder of the section.
Since
r0 Ax0 − b Ax0 − Ax ∗ A(x0 − x ∗ ),
we have that
xk+1 − x ∗ x0 + Pk∗ (A)r0 − x ∗ [I + Pk∗ (A)A](x0 − x ∗ ).
(5.30)
113
114
CHAPTER 5.
CONJUGATE GRADIENT METHODS
Let 0 < λ1 ≤ λ2 ≤ · · · ≤ λn be the eigenvalues of A, and let v1 , v2 , . . . , vn be the
corresponding orthonormal eigenvectors, so that
n
A
λi vi viT .
i1
Since the eigenvectors span the whole space IRn , we can write
x0 − x ∗ n
ξi vi ,
(5.31)
i1
for some coefficients ξi . It is easy to show that any eigenvector of A is also an eigenvector
of Pk (A) for any polynomial Pk . For our particular matrix A and its eigenvalues λi and
eigenvectors vi , we have
Pk (A)vi Pk (λi )vi ,
i 1, 2, . . . , n.
By substituting (5.31) into (5.30) we have
xk+1 − x ∗ n
[1 + λi Pk∗ (λi )]ξi vi .
i1
By using the fact that z2A z T Az n
i1
xk+1 − x ∗ 2A n
λi (viT z)2 , we have
λi [1 + λi Pk∗ (λi )]2 ξi2 .
(5.32)
i1
Since the polynomial Pk∗ generated by the CG method is optimal with respect to this norm,
we have
xk+1 − x ∗ 2A min
n
Pk
λi [1 + λi Pk (λi )]2 ξi2 .
i1
By extracting the largest of the terms [1 + λi Pk (λi )]2 from this expression, we obtain that
⎛
xk+1 − x ∗ 2A ≤ min max [1 + λi Pk (λi )]2 ⎝
Pk 1≤i≤n
n
⎞
λ j ξ 2j ⎠
j1
min max [1 + λi Pk (λi )] x0 − x ∗ 2A ,
2
Pk 1≤i≤n
where we have used the fact that x0 − x ∗ 2A n
j1
λ j ξ 2j .
(5.33)
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
The expression (5.33) allows us to quantify the convergence rate of the CG method
by estimating the nonnegative scalar quantity
min max [1 + λi Pk (λi )]2 .
(5.34)
Pk 1≤i≤n
In other words, we search for a polynomial Pk that makes this expression as small as possible.
In some practical cases, we can find this polynomial explicitly and draw some interesting
conclusions about the properties of the CG method. The following result is an example.
Theorem 5.4.
If A has only r distinct eigenvalues, then the CG iteration will terminate at the solution
in at most r iterations.
PROOF. Suppose that the eigenvalues λ1 , λ2 , . . . , λn take on the r distinct values τ1 < τ2 <
· · · < τr . We define a polynomial Q r (λ) by
Q r (λ) (−1)r
(λ − τ1 )(λ − τ2 ) · · · (λ − τr ),
τ1 τ2 · · · τr
and note that Q r (λi ) 0 for i 1, 2, . . . , n and Q r (0) 1. From the latter observation,
we deduce that Q r (λ) − 1 is a polynomial of degree r with a root at λ 0, so by polynomial
division, the function P̄r −1 defined by
P̄r −1 (λ) (Q r (λ) − 1)/λ
is a polynomial of degree r − 1. By setting k r − 1 in (5.34), we have
0 ≤ min max [1 + λi Pr −1 (λi )]2 ≤ max [1 + λi P̄r −1 (λi )]2 max Q r2 (λi ) 0.
Pr −1 1≤i≤n
1≤i≤n
1≤i≤n
Hence, the constant in (5.34) is zero for the value k r − 1, so we have by substituting into
(5.33) that xr − x ∗ 2A 0, and therefore xr x ∗ , as claimed.
By using similar reasoning, Luenberger [195] establishes the following estimate, which
gives a useful characterization of the behavior of the CG method.
Theorem 5.5.
If A has eigenvalues λ1 ≤ λ2 ≤ · · · ≤ λn , we have that
xk+1 −
x ∗ 2A
≤
λn−k − λ1
λn−k + λ1
2
x0 − x ∗ 2A .
(5.35)
115
116
CHAPTER 5.
CONJUGATE GRADIENT METHODS
λ1
0
Figure 5.3
|
1
λ n−m
λ n−m+1
λn
Two clusters of eigenvalues.
Without giving details of the proof, we describe how this result is obtained from (5.33). One
selects a polynomial P̄k of degree k such that the polynomial Q k+1 (λ) 1 + λ P̄k (λ) has
roots at the k largest eigenvalues λn , λn−1 , . . . , λn−k+1 , as well as at the midpoint between
λ1 and λn−k . It can be shown that the maximum value attained by Q k+1 on the remaining
eigenvalues λ1 , λ2 , . . . , λn−k is precisely (λn−k − λ1 )/(λn−k + λ1 ).
We now illustrate how Theorem 5.5 can be used to predict the behavior of the CG
method on specific problems. Suppose we have the situation plotted in Figure 5.3, where
the eigenvalues of A consist of m large values, with the remaining n − m smaller eigenvalues
clustered around 1. If we define λn−m − λ1 , Theorem 5.5 tells us that after m + 1 steps
of the conjugate gradient algorithm, we have
xm+1 − x ∗ A ≈ x0 − x ∗ A .
For a small value of , we conclude that the CG iterates will provide a good estimate of the
solution after only m + 1 steps.
Figure 5.4 shows the behavior of CG on a problem of this type, which has five large
eigenvalues with all the smaller eigenvalues clustered between 0.95 and 1.05, and compares
this behavior with that of CG on a problem in which the eigenvalues satisfy some random
distribution. In both cases, we plot the log of φ after each iteration.
For the problem with clustered eigenvalues, Theorem 5.5 predicts a sharp decrease in
the error measure at iteration 6. Note, however, that this decrease was achieved one iteration
earlier, illustrating the fact that Theorem 5.5 gives only an upper bound, and that the rate of
convergence can be faster. By contrast, we observe in Figure 5.4 that for the problem with
randomly distributed eigenvalues (dashed line), the convergence rate is slower and more
uniform.
Figure 5.4 illustrates another interesting feature: After one more iteration (a total
of seven) on the problem with clustered eigenvalues, the error measure drops sharply. An
extension of the arguments leading to Theorem 5.4 explains this behavior. It is almost
true to say that the matrix A has just six distinct eigenvalues: the five large eigenvalues
and 1. Then we would expect the error measure to be zero after six iterations. Because the
eigenvalues near 1 are slightly spread out, however, the error does not become very small until
iteration 7.
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
log(||x-x*||A2)
5
clustered eigenvalues
0
uniformly distributed
eigenvalues
-5
-10
1
2
3
4
5
6
7
iteration
Figure 5.4 Performance of the conjugate gradient method on (a) a problem in which
five of the eigenvalues are large and the remainder are clustered near 1, and (b) a matrix
with uniformly distributed eigenvalues.
To state this claim more precisely, it is generally true that if the eigenvalues occur in r
distinct clusters, the CG iterates will approximately solve the problem in about r steps (see
[136]). This result can be proved by constructing a polynomial P̄r −1 such that (1+λ P̄r −1 (λ))
has zeros inside each of the clusters. This polynomial may not vanish at the eigenvalues λi ,
i 1, 2, . . . , n, but its value will be small at these points, so the constant defined in (5.34)
will be small for k ≥ r − 1. We illustrate this behavior in Figure 5.5, which shows the
performance of CG on a matrix of dimension n 14 that has four clusters of eigenvalues:
single eigenvalues at 140 and 120, a cluster of 10 eigenvalues very close to 10, with the
remaining eigenvalues clustered between 0.95 and 1.05. After four iterations, the error has
decreased significantly. After six iterations, the solution is identified to good accuracy.
Another, more approximate, convergence expression for CG is based on the Euclidean
condition number of A, which is defined by
κ(A) A2 A−1 2 λn /λ1 .
It can be shown that
k
√
κ(A) − 1
xk − x A ≤ 2 √
x0 − x ∗ A .
κ(A) + 1
∗
(5.36)
This bound often gives a large overestimate of the error, but it can be useful in those cases
117
118
CHAPTER 5.
CONJUGATE GRADIENT METHODS
log(||x-x*||A2)
5
0
-5
-10
1
2
3
4
5
6
7
iteration
Figure 5.5 Performance of the conjugate gradient method on a matrix in which the
eigenvalues occur in four distinct clusters.
where the only information we have about A is estimates of the extreme eigenvalues λ1
and λn . This bound should be compared with that of the steepest descent method given by
(3.29), which is identical in form but which depends on the condition number κ(A), and
√
not on its square root κ(A).
PRECONDITIONING
We can accelerate the conjugate gradient method by transforming the linear system
to improve the eigenvalue distribution of A. The key to this process, which is known as
preconditioning, is a change of variables from x to x̂ via a nonsingular matrix C, that is,
x̂ C x.
(5.37)
The quadratic φ defined by (5.2) is transformed accordingly to
φ̂(x̂) 12 x̂ T (C −T AC −1 )x̂ − (C −T b)T x̂.
(5.38)
If we use Algorithm 5.2 to minimize φ̂ or, equivalently, to solve the linear system
(C −T AC −1 )x̂ C −T b,
then the convergence rate will depend on the eigenvalues of the matrix C −T AC −1 rather
than those of A. Therefore, we aim to choose C such that the eigenvalues of C −T AC −1
5.1.
THE LINEAR CONJUGATE GRADIENT METHOD
are more favorable for the convergence theory discussed above. We can try to choose C
such that the condition number of C −T AC −1 is much smaller than the original condition
number of A, for instance, so that the constant in (5.36) is smaller. We could also try to
choose C such that the eigenvalues of C −T AC −1 are clustered, which by the discussion of
the previous section ensures that the number of iterates needed to find a good approximate
solution is not much larger than the number of clusters.
It is not necessary to carry out the transformation (5.37) explicitly. Rather, we can
apply Algorithm 5.2 to the problem (5.38), in terms of the variables x̂, and then invert the
transformations to reexpress all the equations in terms of x. This process of derivation results
in Algorithm 5.3 (Preconditioned Conjugate Gradient), which we now define. It happens
that Algorithm 5.3 does not make use of C explicitly, but rather the matrix M C T C,
which is symmetric and positive definite by construction.
Algorithm 5.3 (Preconditioned CG).
Given x0 , preconditioner M;
Set r0 ← Ax0 − b;
Solve M y0 r0 for y0 ;
Set p0 −y0 , k ← 0;
while rk 0
rkT yk
;
pkT Apk
← xk + αk pk ;
αk ←
xk+1
rk+1 ← rk + αk Apk ;
(5.39a)
(5.39b)
(5.39c)
Solve M yk+1 rk+1 ;
r T yk+1
βk+1 ← k+1T
;
rk yk
pk+1 ← −yk+1 + βk+1 pk ;
(5.39d)
k ← k + 1;
(5.39g)
(5.39e)
(5.39f)
end (while)
If we set M I in Algorithm 5.3, we recover the standard CG method, Algorithm 5.2.
The properties of Algorithm 5.2 generalize to this case in interesting ways. In particular, the
orthogonality property (5.16) of the successive residuals becomes
riT M −1r j 0
for all i j.
(5.40)
119
120
CHAPTER 5.
CONJUGATE GRADIENT METHODS
In terms of computational effort, the main difference between the preconditioned
and unpreconditioned CG methods is the need to solve systems of the form M y r (step
(5.39d)).
PRACTICAL PRECONDITIONERS
No single preconditioning strategy is “best” for all conceivable types of matrices:
The tradeoff between various objectives—effectiveness of M, inexpensive computation and
storage of M, inexpensive solution of M y r —varies from problem to problem.
Good preconditioning strategies have been devised for specific types of matrices, in
particular, those arising from discretizations of partial differential equations (PDEs). Often,
the preconditioner is defined in such a way that the system M y r amounts to a simplified
version of the original system Ax b. In the case of a PDE, M y r could represent
a coarser discretization of the underlying continuous problem than Ax b. As in many
other areas of optimization and numerical analysis, knowledge about the structure and
origin of a problem (in this case, knowledge that the system Ax b is a finite-dimensional
representation of a PDE) is the key to devising effective techniques for solving the problem.
General-purpose preconditioners have also been proposed, but their success varies
greatly from problem to problem. The most important strategies of this type include symmetric successive overrelaxation (SSOR), incomplete Cholesky, and banded preconditioners.
(See [272], [136], and [72] for discussions of these techniques.) Incomplete Cholesky is probably the most effective in general. The basic idea is simple: We follow the Cholesky procedure,
but instead of computing the exact Cholesky factor L that satisfies A L L T , we compute
an approximate factor L̃ that is sparser than L. (Usually, we require L̃ to be no denser, or
not much denser, than the lower triangle of the original matrix A.) We then have A ≈ L̃ L̃ T ,
and by choosing C L̃ T , we obtain M L̃ L̃ T and
C −T AC −1 L̃ −1 A L̃ −T ≈ I,
so the eigenvalue distribution of C −T AC −1 is favorable. We do not compute M explicitly,
but rather store the factor L̃ and solve the system M y r by performing two triangular
substitutions with L̃. Because the sparsity of L̃ is similar to that of A, the cost of solving
M y r is similar to the cost of computing the matrix–vector product Ap.
There are several possible pitfalls in the incomplete Cholesky approach. One is that
the resulting matrix may not be (sufficiently) positive definite, and in this case one may need
to increase the values of the diagonal elements to ensure that a value for L̃ can be found.
Numerical instability or breakdown can occur during the incomplete factorization because
of the sparsity conditions we impose on the factor L̃. This difficulty can be remedied by
allowing additional fill-in in L̃, but the denser factor will be more expensive to compute and
to apply at each iteration.
5.2.
5.2
NONLINEAR CONJUGATE GRADIENT METHODS
NONLINEAR CONJUGATE GRADIENT METHODS
We have noted that the CG method, Algorithm 5.2, can be viewed as a minimization
algorithm for the convex quadratic function φ defined by (5.2). It is natural to ask whether
we can adapt the approach to minimize general convex functions, or even general nonlinear
functions f . In fact, as we show in this section, nonlinear variants of the conjugate gradient
are well studied and have proved to be quite successful in practice.
THE FLETCHER–REEVES METHOD
Fletcher and Reeves [107] showed how to extend the conjugate gradient method to
nonlinear functions by making two simple changes in Algorithm 5.2. First, in place of
the formula (5.24a) for the step length αk (which minimizes φ along the search direction
pk ), we need to perform a line search that identifies an approximate minimum of the
nonlinear function f along pk . Second, the residual r , which is simply the gradient of φ in
Algorithm 5.2 (see (5.3)), must be replaced by the gradient of the nonlinear objective f .
These changes give rise to the following algorithm for nonlinear optimization.
Algorithm 5.4 (FR).
Given x0 ;
Evaluate f 0 f (x0 ), ∇ f 0 ∇ f (x0 );
Set p0 ← −∇ f 0 , k ← 0;
while ∇ f k 0
Compute αk and set xk+1 xk + αk pk ;
Evaluate ∇ f k+1 ;
FR
βk+1
←
pk+1
T
∇ f k+1
∇ f k+1
;
∇ f kT ∇ f k
FR
← −∇ f k+1 + βk+1
pk ;
k ← k + 1;
(5.41a)
(5.41b)
(5.41c)
end (while)
If we choose f to be a strongly convex quadratic and αk to be the exact minimizer, this
algorithm reduces to the linear conjugate gradient method, Algorithm 5.2. Algorithm 5.4
is appealing for large nonlinear optimization problems because each iteration requires only
evaluation of the objective function and its gradient. No matrix operations are required for
the step computation, and just a few vectors of storage are required.
To make the specification of Algorithm 5.4 complete, we need to be more precise
about the choice of line search parameter αk . Because of the second term in (5.41b), the
search direction pk may fail to be a descent direction unless αk satisfies certain conditions.
121
122
CHAPTER 5.
CONJUGATE GRADIENT METHODS
By taking the inner product of (5.41b) (with k replacing k + 1) with the gradient vector
∇ f k , we obtain
∇ f kT pk −∇ f k 2 + βkFR ∇ f kT pk−1 .
(5.42)
If the line search is exact, so that αk−1 is a local minimizer of f along the direction pk−1 ,
we have that ∇ f kT pk−1 0. In this case we have from (5.42) that ∇ f kT pk < 0, so that pk is
indeed a descent direction. If the line search is not exact, however, the second term in (5.42)
may dominate the first term, and we may have ∇ f kT pk > 0, implying that pk is actually a
direction of ascent. Fortunately, we can avoid this situation by requiring the step length αk
to satisfy the strong Wolfe conditions, which we restate here:
f (xk + αk pk ) ≤ f (xk ) + c1 αk ∇ f kT pk ,
|∇ f (xk + αk pk ) pk | ≤ −c2 ∇
T
f kT
pk ,
(5.43a)
(5.43b)
where 0 < c1 < c2 < 12 . (Note that we impose c2 < 12 here, in place of the looser condition
c2 < 1 that was used in the earlier statement (3.7).) By applying Lemma 5.6 below, we can
show that condition (5.43b) implies that (5.42) is negative, and we conclude that any line
search procedure that yields an αk satisfying (5.43) will ensure that all directions pk are
descent directions for the function f .
THE POLAK–RIBIÈRE METHOD AND VARIANTS
There are many variants of the Fletcher–Reeves method that differ from each other
mainly in the choice of the parameter βk . An important variant, proposed by Polak and
Ribière, defines this parameter as follows:
PR
βk+1
T
(∇ f k+1 − ∇ f k )
∇ f k+1
.
∇ f k 2
(5.44)
We refer to the algorithm in which (5.44) replaces (5.41a) as Algorithm PR. It is identical to
Algorithm FR when f is a strongly convex quadratic function and the line search is exact,
PR
FR
βk+1
. When applied
since by (5.16) the gradients are mutually orthogonal, and so βk+1
to general nonlinear functions with inexact line searches, however, the behavior of the two
algorithms differs markedly. Numerical experience indicates that Algorithm PR tends to be
the more robust and efficient of the two.
A surprising fact about Algorithm PR is that the strong Wolfe conditions (5.43) do
not guarantee that pk is always a descent direction. If we define the β parameter as
+
PR
βk+1
max{βk+1
, 0},
(5.45)
5.2.
NONLINEAR CONJUGATE GRADIENT METHODS
giving rise to an algorithm we call Algorithm PR+, then a simple adaptation of the strong
Wolfe conditions ensures that the descent property holds.
There are many other choices for βk+1 that coincide with the Fletcher–Reeves formula
FR
βk+1 in the case where the objective is quadratic and the line search is exact. The Hestenes–
Stiefel formula, which defines
HS
βk+1
T
(∇ f k+1 − ∇ f k )
∇ f k+1
,
(∇ f k+1 − ∇ f k )T pk
(5.46)
gives rise to an algorithm (called Algorithm HS) that is similar to Algorithm PR, both in
terms of its theoretical convergence properties and in its practical performance. Formula
(5.46) can be derived by demanding that consecutive search directions be conjugate with
respect to the average Hessian over the line segment [xk , xk+1 ], which is defined as
1
Ḡ k ≡
[∇ 2 f (xk + τ αk pk )]dτ.
0
Recalling from Taylor’s theorem (Theorem 2.1) that ∇ f k+1 ∇ f k + αk Ḡ k pk , we see that
T
Ḡ k pk 0
for any direction of the form pk+1 −∇ f k+1 + βk+1 pk , the condition pk+1
requires βk+1 to be given by (5.46).
Later, we see that it is possible to guarantee global convergence for any parameter βk
satisfying the bound
|βk | ≤ βkFR ,
(5.47)
for all k ≥ 2. This fact suggests the following modification of the PR method, which has
performed well on some applications. For all k ≥ 2 let
⎧
FR
⎪
⎨ −βk
βk βkPR
⎪
⎩
βkFR
if
βkPR
<
−βkFR
if
|βkPR |
≤
βkFR
if
βk
>
βk .
PR
(5.48)
FR
The algorithm based on this strategy will be denoted by FR-PR.
Other variants of the CG method have recently been proposed. Two choices for βk+1
that possess attractive theoretical and computational properties are
βk+1 ∇ f k+1 2
(∇ f k+1 − ∇ f k )T pk
(5.49)
(see [85]) and
βk+1 ŷk − 2 pk
ŷk 2
ŷkT pk
T
∇ f k+1
,
ŷkT pk
with
ŷk ∇ f k+1 − ∇ f k
(5.50)
123
124
CHAPTER 5.
CONJUGATE GRADIENT METHODS
(see [161]). These two choices guarantee that pk is a descent direction, provided the
steplength αk satisfies the Wolfe conditions. The CG algorithms based on (5.49) or (5.50)
appear to be competitive with the Polak–Ribière method.
QUADRATIC TERMINATION AND RESTARTS
Implementations of nonlinear conjugate gradient methods usually preserve their
close connections with the linear conjugate gradient method. Usually, a quadratic (or cubic)
interpolation along the search direction pk is incorporated into the line search procedure; see
Chapter 3. This feature guarantees that when f is a strictly convex quadratic, the step length
αk is chosen to be the exact one-dimensional minimizer, so that the nonlinear conjugate
gradient method reduces to the linear method, Algorithm 5.2.
Another modification that is often used in nonlinear conjugate gradient procedures
is to restart the iteration at every n steps by setting βk 0 in (5.41a), that is, by taking
a steepest descent step. Restarting serves to periodically refresh the algorithm, erasing old
information that may not be beneficial. We can even prove a strong theoretical result about
restarting: It leads to n-step quadratic convergence, that is,
xk+n − x O xk − x ∗ 2 .
(5.51)
After a little thought, this result is not so surprising. Consider a function f that is strongly
convex quadratic in a neighborhood of the solution, but is nonquadratic everywhere else.
Assuming that the algorithm is converging to the solution in question, the iterates will
eventually enter the quadratic region. At some point, the algorithm will be restarted in that
region, and from that point onward, its behavior will simply be that of the linear conjugate
gradient method, Algorithm 5.2. In particular, finite termination will occur within n steps
of the restart. The restart is important, because the finite-termination property and other
appealing properties of Algorithm 5.2 hold only when its initial search direction p0 is equal
to the negative gradient.
Even if the function f is not exactly quadratic in the region of a solution, Taylor’s
theorem (Theorem 2.1) implies that it can still be approximated quite closely by a quadratic,
provided that it is smooth. Therefore, while we would not expect termination in n steps
after the restart, it is not surprising that substantial progress is made toward the solution, as
indicated by the expression (5.51).
Though the result (5.51) is interesting from a theoretical viewpoint, it may not be
relevant in a practical context, because nonlinear conjugate gradient methods can be recommended only for solving problems with large n. Restarts may never occur in such problems
because an approximate solution may be located in fewer than n steps. Hence, nonlinear
CG method are sometimes implemented without restarts, or else they include strategies for
restarting that are based on considerations other than iteration counts. The most popular
restart strategy makes use of the observation (5.16), which is that the gradients are mutually
orthogonal when f is a quadratic function. A restart is performed whenever two consecutive
5.2.
NONLINEAR CONJUGATE GRADIENT METHODS
gradients are far from orthogonal, as measured by the test
|∇ f kT ∇ f k−1 |
≥ ν,
∇ f k 2
(5.52)
where a typical value for the parameter ν is 0.1.
We could also think of formula (5.45) as a restarting strategy, because pk+1 will revert
to the steepest descent direction whenever βkPR is negative. In contrast to (5.52), these restarts
are rather infrequent because βkPR is positive most of the time.
BEHAVIOR OF THE FLETCHER–REEVES METHOD
We now investigate the Fletcher–Reeves algorithm, Algorithm 5.4, a little more closely,
proving that it is globally convergent and explaining some of its observed inefficiencies.
The following result gives conditions on the line search under which all search directions are descent directions. It assumes that the level set L {x : f (x) ≤ f (x0 )} is bounded
and that f is twice continuously differentiable, so that we have from Lemma 3.1 that there
exists a step length αk satisfying the strong Wolfe conditions.
Lemma 5.6.
Suppose that Algorithm 5.4 is implemented with a step length αk that satisfies the strong
Wolfe conditions (5.43) with 0 < c2 < 12 . Then the method generates descent directions pk
that satisfy the following inequalities:
−
∇ f kT pk
2c2 − 1
1
≤
≤
,
1 − c2
∇ f k 2
1 − c2
for all k 0, 1, . . . .
(5.53)
def
PROOF. Note first that the function t(ξ ) (2ξ − 1)/(1 − ξ ) is monotonically increasing
on the interval [0, 12 ] and that t(0) −1 and t( 21 ) 0. Hence, because of c2 ∈ (0, 12 ), we
have
−1<
2c2 − 1
< 0.
1 − c2
(5.54)
The descent condition ∇ f kT pk < 0 follows immediately once we establish (5.53).
The proof is by induction. For k 0, the middle term in (5.53) is −1, so by using
(5.54), we see that both inequalities in (5.53) are satisfied. Next, assume that (5.53) holds
for some k ≥ 1. From (5.41b) and (5.41a) we have
T
T
T
pk+1
∇ f k+1
pk
pk
∇ f k+1
∇ f k+1
−1
+
β
−1
+
.
k+1
2
2
∇ f k+1 ∇ f k+1 ∇ f k 2
(5.55)
125
126
CHAPTER 5.
CONJUGATE GRADIENT METHODS
By using the line search condition (5.43b), we have
T
pk | ≤ −c2 ∇ f kT pk ,
|∇ f k+1
so by combining with (5.55) and recalling (5.41a), we obtain
−1 + c2
T
pk+1
∇ f k+1
∇ f kT pk
∇ f kT pk
≤
≤
−1
−
c
.
2
∇ f k 2
∇ f k+1 2
∇ f k 2
Substituting for the term ∇ f kT pk /∇ f k 2 from the left-hand-side of the induction
hypothesis (5.53), we obtain
−1 −
T
pk+1
∇ f k+1
c2
c2
≤
≤ −1 +
,
2
1 − c2
∇ f k+1 1 − c2
which shows that (5.53) holds for k + 1 as well.
This result used only the second strong Wolfe condition (5.43b); the first Wolfe
condition (5.43a) will be needed in the next section to establish global convergence. The
bounds on ∇ f kT pk in (5.53) impose a limit on how fast the norms of the steps pk can
grow, and they will play a crucial role in the convergence analysis given below.
Lemma 5.6 can also be used to explain a weakness of the Fletcher–Reeves method.
We will argue that if the method generates a bad direction and a tiny step, then the next
direction and next step are also likely to be poor. As in Chapter 3, we let θk denote the angle
between pk and the steepest descent direction −∇ f k , defined by
cos θk −∇ f kT pk
.
∇ f k pk (5.56)
Suppose that pk is a poor search direction, in the sense that it makes an angle of nearly 90◦
with −∇ f k , that is, cos θk ≈ 0. By multiplying both sides of (5.53) by ∇ f k / pk and
using (5.56), we obtain
1 ∇ f k 1 − 2c2 ∇ f k ≤ cos θk ≤
,
1 − c2 pk 1 − c2 pk for all k 0, 1, . . ..
(5.57)
From these inequalities, we deduce that cos θk ≈ 0 if and only if
∇ f k pk .
Since pk is almost orthogonal to the gradient, it is likely that the step from xk to xk+1 is tiny,
that is, xk+1 ≈ xk . If so, we have ∇ f k+1 ≈ ∇ f k , and therefore
FR
≈ 1,
βk+1
(5.58)
5.2.
NONLINEAR CONJUGATE GRADIENT METHODS
by the definition (5.41a). By using this approximation together with ∇ f k+1 ≈ ∇ f k pk in (5.41b), we conclude that
pk+1 ≈ pk ,
so the new search direction will improve little (if at all) on the previous one. It follows that
if the condition cos θk ≈ 0 holds at some iteration k and if the subsequent step is small, a
long sequence of unproductive iterates will follow.
The Polak–Ribière method behaves quite differently in these circumstances. If, as in
the previous paragraph, the search direction pk satisfies cos θk ≈ 0 for some k, and if the
PR
≈ 0.
subsequent step is small, it follows by substituting ∇ f k ≈ ∇ f k+1 into (5.44) that βk+1
From the formula (5.41b), we find that the new search direction pk+1 will be close to the
steepest descent direction −∇ f k+1 , and cos θk+1 will be close to 1. Therefore, Algorithm PR
essentially performs a restart after it encounters a bad direction. The same argument can
be applied to Algorithms PR+ and HS. For the FR-PR variant, defined by (5.48), we have
FR
PR
PR
≈ 1, and βk+1
≈ 0. The formula (5.48) thus sets βk+1 βk+1
, as
noted already that βk+1
desired. Thus, the modification (5.48) seems to avoid the inefficiencies of the FR method,
while falling back on this method for global convergence.
The undesirable behavior of the Fletcher–Reeves method predicted by the arguments
given above can be observed in practice. For example, the paper [123] describes a problem
with n 100 in which cos θk is of order 10−2 for hundreds of iterations and the steps
xk − xk−1 are of order 10−2 . Algorithm FR requires thousands of iterations to solve this
problem, while Algorithm PR requires just 37 iterations. In this example, the Fletcher–
Reeves method performs much better if it is periodically restarted along the steepest descent
direction, since each restart terminates the cycle of bad steps. In general, Algorithm FR
should not be implemented without some kind of restart strategy.
GLOBAL CONVERGENCE
Unlike the linear conjugate gradient method, whose convergence properties are well
understood and which is known to be optimal as described above, nonlinear conjugate
gradient methods possess surprising, sometimes bizarre, convergence properties. We now
present a few of the main results known for the Fletcher–Reeves and Polak–Ribière methods
using practical line searches.
For the purposes of this section, we make the following (nonrestrictive) assumptions
on the objective function.
Assumptions 5.1.
(i) The level set L : {x | f (x) ≤ f (x0 )} is bounded;
(ii) In some open neighborhood N of L, the objective function f is Lipschitz continuously
differentiable.
127
128
CHAPTER 5.
CONJUGATE GRADIENT METHODS
These assumptions imply that there is a constant γ̄ such that
∇ f (x) ≤ γ̄ , for all x ∈ L.
(5.59)
Our main analytical tool in this section is Zoutendijk’s theorem—Theorem 3.2 in
Chapter 3. It states, that under Assumptions 5.1, any line search iteration of the form
xk+1 xk + αk pk , where pk is a descent direction and αk satisfies the Wolfe conditions
(5.43) gives the limit
∞
cos2 θk ∇ f k 2 < ∞.
(5.60)
k0
We can use this result to prove global convergence for algorithms that are periodically
restarted by setting βk 0. If k1 , k2 , and so on denote the iterations on which restarts occur,
we have from (5.60) that
∇ f k 2 < ∞.
(5.61)
kk1 ,k2 ,...
If we allow no more than n̄ iterations between restarts, the sequence {k j }∞
j1 is infinite,
and from (5.61) we have that lim j→∞ ∇ f k j 0. That is, a subsequence of gradients
approaches zero, or equivalently,
lim inf ∇ f k 0.
k→∞
(5.62)
This result applies equally to restarted versions of all the algorithms discussed in this chapter.
It is more interesting, however, to study the global convergence of unrestarted conjugate
gradient methods, because for large problems (say n ≥ 1000) we expect to find a solution in
many fewer than n iterations—the first point at which a regular restart would take place. Our
study of large sequences of unrestarted conjugate gradient iterations reveals some surprising
patterns in their behavior.
We can build on Lemma 5.6 and Zoutendijk’s result (5.60) to prove a global convergence result for the Fletcher–Reeves method. While we cannot show that the limit of the
sequence of gradients {∇ f k } is zero, the following result shows that this sequence is not
bounded away from zero.
Theorem 5.7 (Al-Baali [3]).
Suppose that Assumptions 5.1 hold, and that Algorithm 5.4 is implemented with a line
search that satisfies the strong Wolfe conditions (5.43), with 0 < c1 < c2 < 12 . Then
lim inf ∇ f k 0.
k→∞
(5.63)
5.2.
NONLINEAR CONJUGATE GRADIENT METHODS
PROOF. The proof is by contradiction. It assumes that the opposite of (5.63) holds, that is,
there is a constant γ > 0 such that
∇ f k ≥ γ ,
(5.64)
for all k sufficiently large. By substituting the left inequality of (5.57) into Zoutendijk’s
condition (5.60), we obtain
∞
∇ f k 4
< ∞.
pk 2
k0
(5.65)
By using (5.43b) and (5.53), we obtain that
T
|∇ f kT pk−1 | ≤ −c2 ∇ f k−1
pk−1 ≤
c2
∇ f k−1 2 .
1 − c2
(5.66)
Thus, from (5.41b) and recalling the definition (5.41a) of βkFR we obtain
pk 2 ≤ ∇ f k 2 + 2βkFR |∇ f kT pk−1 | + (βkFR )2 pk−1 2
2c2 FR
≤ ∇ f k 2 +
β ∇ f k−1 2 + (βkFR )2 pk−1 2
1 − c2 k
1 + c2
∇ f k 2 + (βkFR )2 pk−1 2 .
1 − c2
def
Applying this relation repeatedly, and defining c3 (1 + c2 )/(1 − c2 ) ≥ 1, we have
FR
pk 2 ≤ c3 ∇ f k 2 + (βkFR )2 (c3 ∇ f k−1 2 + (βk−1
)2 (c3 ∇ f k−2 2 +
· · · + (β1FR )2 p0 2 )) · · ·)
c3 ∇ f k 4
k
∇ f j −2 ,
(5.67)
j0
where we used the facts that
FR
FR
(βkFR )2 (βk−1
)2 · · · (βk−i
)2 ∇ f k 4
∇ f k−i−1 4
and p0 −∇ f 0 . By using the bounds (5.59) and (5.64) in (5.67), we obtain
pk 2 ≤
c3 γ̄ 4
k,
γ2
(5.68)
129
130
CHAPTER 5.
CONJUGATE GRADIENT METHODS
which implies that
∞
k1
∞
1
1
,
≥ γ4
2
pk k
k1
(5.69)
for some positive constant γ4 .
On the other hand, from (5.64) and (5.65), we have that
∞
k1
1
< ∞.
pk 2
(5.70)
However, if we combine this inequality with (5.69), we obtain that ∞
k1 1/k < ∞, which
is not true. Hence, (5.64) does not hold, and the claim (5.63) is proved.
This global convergence result can be extended to any choice of βk satisfying (5.47),
and in particular to the FR-PR method given by (5.48).
In general, if we can show that there exist constants c4 , c5 > 0 such that
cos θk ≥ c4
∇ f k ,
pk ∇ f k ≥ c5 > 0,
pk k 1, 2, . . . ,
it follows from (5.60) that
lim ∇ f k 0.
k→∞
In fact, this result can be established for the Polak–Ribière method under the assumption
that f is strongly convex and that an exact line search is used.
For general (nonconvex) functions, however, is it not possible to prove a result like
Theorem 5.7 for Algorithm PR. This fact is unexpected, since the Polak–Ribière method
performs better in practice than the Fletcher–Reeves method. The following surprising result
shows that the Polak–Ribière method can cycle infinitely without approaching a solution
point, even if an ideal line search is used. (By “ideal” we mean that line search returns a
value αk that is the first positive stationary point for the function t(α) f (xk + αpk ).)
Theorem 5.8.
Consider the Polak–Ribière method method (5.44) with an ideal line search. There exists
a twice continuously differentiable objective function f : IR3 → IR and a starting point x0 ∈ IR3
such that the sequence of gradients {∇ f k } is bounded away from zero.
The proof of this result, given in [253], is quite complex. It demonstrates the existence
of the desired objective function without actually constructing this function explicitly. The
result is interesting, since the step length assumed in the proof—the first stationary point—
may be accepted by any of the practical line search algorithms currently in use. The proof
5.2.
NONLINEAR CONJUGATE GRADIENT METHODS
of Theorem 5.8 requires that some consecutive search directions become almost negatives
of each other. In the case of ideal line searches, this happens only if βk < 0, so the analysis
suggests Algorithm PR+ (see (5.45)), in which we reset βk to zero whenever it becomes
negative. We mentioned earlier that a line search strategy based on a slight modification of
the Wolfe conditions guarantees that all search directions generated by Algorithm PR+ are
descent directions. Using these facts, it is possible to a prove global convergence result like
Theorem 5.7 for Algorithm PR+. An attractive property of the formulae (5.49), (5.50) is
that global convergence can be established without introducing any modification to a line
search based on the Wolfe conditions.
NUMERICAL PERFORMANCE
Table 5.1 illustrates the performance of Algorithms FR, PR, and PR+ without restarts.
For these tests, the parameters in the strong Wolfe conditions (5.43) were chosen to be
c1 10−4 and c2 0.1. The iterations were terminated when
∇ f k ∞ < 10−5 (1 + | f k |).
If this condition was not satisfied after 10,000 iterations, we declare failure (indicated by a
∗ in the table).
The final column, headed “mod,” indicates the number of iterations of Algorithm PR+
for which the adjustment (5.45) was needed to ensure that βkPR ≥ 0. Algorithm FR on problem
GENROS takes very short steps far from the solution that lead to tiny improvements in the
objective function, and convergence was not achieved within the maximum number of
iterations.
The Polak–Ribière algorithm, or its variation PR+, are not always more efficient
than Algorithm FR, and it has the slight disadvantage of requiring one more vector of
storage. Nevertheless, we recommend that users choose Algorithm PR, PR+ or FR-PR, or
the methods based on (5.49) and (5.50).
Table 5.1 Iterations and function/gradient evaluations required by three
nonlinear conjugate gradient methods on a set of test problems; see [123]
Problem
CALCVAR3
GENROS
XPOWSING
TRIDIA1
MSQRT1
XPOWELL
TRIGON
n
200
500
1000
1000
1000
1000
1000
Alg FR
it/f-g
Alg PR
it/f-g
2808/5617
∗
533/1102
264/531
422/849
568/1175
231/467
2631/5263
1068/2151
212/473
262/527
113/231
212/473
40/92
Alg PR+
it/f-g
mod
2631/5263
1067/2149
97/229
262/527
113/231
97/229
40/92
0
1
3
0
0
3
0
131
132
CHAPTER 5.
CONJUGATE GRADIENT METHODS
NOTES AND REFERENCES
The conjugate gradient method was developed in the 1950s by Hestenes and
Stiefel [168] as an alternative to factorization methods for finding solutions of symmetric positive definite systems. It was not until some years later, in one of the most important
developments in sparse linear algebra, that this method came to be viewed as an iterative
method that could give good approximate solutions to systems in many fewer than n steps.
Our presentation of the linear conjugate gradient method follows that of Luenberger [195].
For a history of the development of the conjugate gradient and Lanczos methods see Golub
and O’Leary [135].
Interestingly enough, the nonlinear conjugate gradient method of Fletcher and
Reeves [107] was proposed after the linear conjugate gradient method had fallen out of
favor, but several years before it was rediscovered as an iterative method for linear systems.
The Polak–Ribière method was introduced in [237], and the example showing that it may
fail to converge on nonconvex problems is given by Powell [253]. Restart procedures are
discussed in Powell [248].
Hager and Zhang [161] report some of the best computational results obtained to date
with a nonlinear CG method. Their implementation is based on formula (5.50) and uses
a high-accuracy line search procedure. The results in Table 5.1 are taken from Gilbert and
Nocedal [123]. This paper also describes a line search that guarantees that Algorithm PR+
always generates descent directions and proves global convergence.
Analysis due to Powell [245] provides further evidence of the inefficiency of the
Fletcher–Reeves method using exact line searches. He shows that if the iterates enter a
region in which the function is the two-dimensional quadratic
f (x) 12 x T x,
then the angle between the gradient ∇ f k and the search direction pk stays constant. Since
this angle can be arbitrarily close to 90◦ , the Fletcher–Reeves method can be slower than
the steepest descent method. The Polak–Ribière method behaves quite differently in these
circumstances: If a very small step is generated, the next search direction tends to the steepest
descent direction, as argued above. This feature prevents a sequence of tiny steps.
The global convergence of nonlinear conjugate gradient methods has received much
attention; see for example Al-Baali [3], Gilbert and Nocedal [123], Dai and Yuan [85], and
Hager and Zhang [161]. For recent surveys on CG methods see Gould et al. [147] and Hager
and Zhang [162].
Most of the theory on the rate of convergence of conjugate gradient methods assumes
that the line search is exact. Crowder and Wolfe [82] show that the rate of convergence
is linear, and show by constructing an example that Q-superlinear convergence is not
achievable. Powell [245] studies the case in which the conjugate gradient method enters a
region where the objective function is quadratic, and shows that either finite termination
occurs or the rate of convergence is linear. Cohen [63] and Burmeister [45] prove n-step
5.2.
NONLINEAR CONJUGATE GRADIENT METHODS
quadratic convergence (5.51) for general objective functions. Ritter [265] shows that in fact,
the rate is superquadratic, that is,
xk+n − x ∗ o(xk − x ∗ 2 ).
Powell [251] gives a slightly better result and performs numerical tests on small problems
to measure the rate observed in practice. He also summarizes rate-of-convergence results
for asymptotically exact line searches, such as those obtained by Baptist and Stoer [11]
and Stoer [282]. Even faster rates of convergence can be established (see Schuller [278],
Ritter [265]), under the assumption that the search directions are uniformly linearly
independent, but this assumption is hard to verify and does not often occur in practice.
Nemirovsky and Yudin [225] devote some attention to the global efficiency of the
Fletcher–Reeves and Polak–Ribière methods with exact line searches. For this purpose they
define a measure of “laboriousness” and an “optimal bound” for it among a certain class
of iterations. They show that on strongly convex problems not only do the Fletcher–Reeves
and Polak–Ribière methods fail to attain the optimal bound, but they may also be slower
than the steepest descent method. Subsequently, Nesterov [225] presented an algorithm that
attains this optimal bound. It is related to PARTAN, the method of parallel tangents (see, for
example, Luenberger [195]). We feel that this approach is unlikely to be effective in practice,
but no conclusive investigation has been carried out, to the best of our knowledge.
✐
EXERCISES
✐
5.1 Implement Algorithm 5.2 and use to it solve linear systems in which A is the
Hilbert matrix, whose elements are Ai, j 1/(i + j − 1). Set the right-hand-side to
b (1, 1, . . . , 1)T and the initial point to x0 0. Try dimensions n 5, 8, 12, 20 and
report the number of iterations required to reduce the residual below 10−6 .
✐
5.2 Show that if the nonzero vectors p0 , p1 , . . . , pl satisfy (5.5), where A is symmetric
and positive definite, then these vectors are linearly independent. (This result implies that
A has at most n conjugate directions.)
✐
5.3 Verify the formula (5.7).
✐
def
5.4 Show that if f (x) is a strictly convex quadratic, then the function h(σ ) f (x0 + σ0 p0 + · · · + σk−1 pk−1 ) also is a strictly convex quadratic in the variable σ (σ0 , σ1 , . . . , σk−1 )T .
✐
5.5 Verify from the formulae (5.14) that (5.17) and (5.18) hold for k 1.
✐
5.6 Show that (5.24d) is equivalent to (5.14d).
133
134
CHAPTER 5.
CONJUGATE GRADIENT METHODS
✐
5.7 Let {λi , vi } i 1, 2, . . . , n be the eigenpairs of the symmetric matrix A. Show
that the eigenvalues and eigenvectors of [I + Pk (A)A]T A[I + Pk (A)A] are λi [1+λi Pk (λi )]2
and vi , respectively.
✐
5.8 Construct matrices with various eigenvalue distributions (clustered and nonclustered) and apply the CG method to them. Comment on whether the behavior can be
explained from Theorem 5.5.
✐
5.9 Derive Algorithm 5.3 by applying the standard CG method in the variables x̂ and
then transforming back into the original variables.
✐
5.10 Verify the modified conjugacy condition (5.40).
✐
5.11 Show that when applied to a quadratic function, with exact line searches, both
the Polak–Ribière formula given by (5.44) and the Hestenes–Stiefel formula given by (5.46)
reduce to the Fletcher–Reeves formula (5.41a).
✐
5.12 Prove that Lemma 5.6 holds for any choice of βk satisfying |βk | ≤ βkFR .
This is page 135
Printer: Opaque this
CHAPTER
6
Quasi-Newton
Methods
In the mid 1950s, W.C. Davidon, a physicist working at Argonne National Laboratory,
was using the coordinate descent method (see Section 9.3) to perform a long optimization
calculation. At that time computers were not very stable, and to Davidon’s frustration,
the computer system would always crash before the calculation was finished. So Davidon
decided to find a way of accelerating the iteration. The algorithm he developed—the first
quasi-Newton algorithm—turned out to be one of the most creative ideas in nonlinear
optimization. It was soon demonstrated by Fletcher and Powell that the new algorithm
was much faster and more reliable than the other existing methods, and this dramatic
136
CHAPTER 6.
QUASI-NEWTON METHODS
advance transformed nonlinear optimization overnight. During the following twenty years,
numerous variants were proposed and hundreds of papers were devoted to their study. An
interesting historical irony is that Davidon’s paper [87] was not accepted for publication; it
remained as a technical report for more than thirty years until it appeared in the first issue
of the SIAM Journal on Optimization in 1991 [88].
Quasi-Newton methods, like steepest descent, require only the gradient of the objective function to be supplied at each iterate. By measuring the changes in gradients, they
construct a model of the objective function that is good enough to produce superlinear
convergence. The improvement over steepest descent is dramatic, especially on difficult
problems. Moreover, since second derivatives are not required, quasi-Newton methods are
sometimes more efficient than Newton’s method. Today, optimization software libraries
contain a variety of quasi-Newton algorithms for solving unconstrained, constrained, and
large-scale optimization problems. In this chapter we discuss quasi-Newton methods for
small and medium-sized problems, and in Chapter 7 we consider their extension to the
large-scale setting.
The development of automatic differentiation techniques has made it possible to use
Newton’s method without requiring users to supply second derivatives; see Chapter 8.
Still, automatic differentiation tools may not be applicable in many situations, and it
may be much more costly to work with second derivatives in automatic differentiation software than with the gradient. For these reasons, quasi-Newton methods remain
appealing.
6.1
THE BFGS METHOD
The most popular quasi-Newton algorithm is the BFGS method, named for its discoverers
Broyden, Fletcher, Goldfarb, and Shanno. In this section we derive this algorithm (and
its close relative, the DFP algorithm) and describe its theoretical properties and practical
implementation.
We begin the derivation by forming the following quadratic model of the objective
function at the current iterate xk :
m k ( p) f k + ∇ f kT p + 12 p T Bk p.
(6.1)
Here Bk is an n × n symmetric positive definite matrix that will be revised or updated at
every iteration. Note that the function value and gradient of this model at p 0 match
f k and ∇ f k , respectively. The minimizer pk of this convex quadratic model, which we can
write explicitly as
pk −Bk−1 ∇ f k ,
(6.2)
6.1.
THE BFGS METHOD
is used as the search direction, and the new iterate is
xk+1 xk + αk pk ,
(6.3)
where the step length αk is chosen to satisfy the Wolfe conditions (3.6). This iteration is
quite similar to the line search Newton method; the key difference is that the approximate
Hessian Bk is used in place of the true Hessian.
Instead of computing Bk afresh at every iteration, Davidon proposed to update it in a
simple manner to account for the curvature measured during the most recent step. Suppose
that we have generated a new iterate xk+1 and wish to construct a new quadratic model, of
the form
T
p + 12 p T Bk+1 p.
m k+1 ( p) f k+1 + ∇ f k+1
What requirements should we impose on Bk+1 , based on the knowledge gained during
the latest step? One reasonable requirement is that the gradient of m k+1 should match the
gradient of the objective function f at the latest two iterates xk and xk+1 . Since ∇m k+1 (0) is
precisely ∇ f k+1 , the second of these conditions is satisfied automatically. The first condition
can be written mathematically as
∇m k+1 (−αk pk ) ∇ f k+1 − αk Bk+1 pk ∇ f k .
By rearranging, we obtain
Bk+1 αk pk ∇ f k+1 − ∇ f k .
(6.4)
To simplify the notation it is useful to define the vectors
sk xk+1 − xk αk pk ,
yk ∇ f k+1 − ∇ f k ,
(6.5)
so that (6.4) becomes
Bk+1 sk yk .
(6.6)
We refer to this formula as the secant equation.
Given the displacement sk and the change of gradients yk , the secant equation requires
that the symmetric positive definite matrix Bk+1 map sk into yk . This will be possible only
if sk and yk satisfy the curvature condition
skT yk > 0,
(6.7)
as is easily seen by premultiplying (6.6) by skT . When f is strongly convex, the inequality (6.7)
will be satisfied for any two points xk and xk+1 (see Exercise 6.1). However, this condition
137
138
CHAPTER 6.
QUASI-NEWTON METHODS
will not always hold for nonconvex functions, and in this case we need to enforce (6.7)
explicitly, by imposing restrictions on the line search procedure that chooses the step length
α. In fact, the condition (6.7) is guaranteed to hold if we impose the Wolfe (3.6) or strong
Wolfe conditions (3.7) on the line search. To verify this claim, we note from (6.5) and (3.6b)
T
sk ≥ c2 ∇ f kT sk , and therefore
that ∇ f k+1
ykT sk ≥ (c2 − 1)αk ∇ f kT pk .
(6.8)
Since c2 < 1 and since pk is a descent direction, the term on the right is positive, and the
curvature condition (6.7) holds.
When the curvature condition is satisfied, the secant equation (6.6) always has a
solution Bk+1 . In fact, it admits an infinite number of solutions, since the n(n + 1)/2
degrees of freedom in a symmetric positive definite matrix exceed the n conditions imposed
by the secant equation. The requirement of positive definiteness imposes n additional
inequalities—all principal minors must be positive—but these conditions do not absorb the
remaining degrees of freedom.
To determine Bk+1 uniquely, we impose the additional condition that among all
symmetric matrices satisfying the secant equation, Bk+1 is, in some sense, closest to the current
matrix Bk . In other words, we solve the problem
min B − Bk (6.9a)
B
B BT ,
subject to
Bsk yk ,
(6.9b)
where sk and yk satisfy (6.7) and Bk is symmetric and positive definite. Different matrix
norms can be used in (6.9a), and each norm gives rise to a different quasi-Newton method.
A norm that allows easy solution of the minimization problem (6.9) and gives rise to a
scale-invariant optimization method is the weighted Frobenius norm
AW ≡ W 1/2 AW 1/2 F ,
(6.10)
n n
2
where · F is defined by C2F i1
j1 ci j . The weight matrix W can be chosen as
any matrix satisfying the relation W yk sk . For concreteness, the reader can assume that
W Ḡ −1
k where Ḡ k is the average Hessian defined by
Ḡ k 1
∇ f (xk + τ αk pk )dτ .
2
(6.11)
0
The property
yk Ḡ k αk pk Ḡ k sk
(6.12)
follows from Taylor’s theorem, Theorem 2.1. With this choice of weighting matrix W , the
6.1.
THE BFGS METHOD
norm (6.10) is non-dimensional, which is a desirable property, since we do not wish the
solution of (6.9) to depend on the units of the problem.
With this weighting matrix and this norm, the unique solution of (6.9) is
(DFP)
Bk+1 I − ρk yk skT Bk I − ρk sk ykT + ρk yk ykT ,
(6.13)
with
ρk 1
ykT sk
.
(6.14)
This formula is called the DFP updating formula, since it is the one originally proposed by
Davidon in 1959, and subsequently studied, implemented, and popularized by Fletcher and
Powell.
The inverse of Bk , which we denote by
Hk Bk−1 ,
is useful in the implementation of the method, since it allows the search direction (6.2)
to be calculated by means of a simple matrix–vector multiplication. Using the Sherman–
Morrison–Woodbury formula (A.28), we can derive the following expression for the update
of the inverse Hessian approximation Hk that corresponds to the DFP update of Bk in (6.13):
(DFP)
Hk+1 Hk −
Hk yk ykT Hk
sk skT
+
.
ykT Hk yk
ykT sk
(6.15)
Note that the last two terms in the right-hand-side of (6.15) are rank-one matrices, so that Hk
undergoes a rank-two modification. It is easy to see that (6.13) is also a rank-two modification
of Bk . This is the fundamental idea of quasi-Newton updating: Instead of recomputing the
approximate Hessians (or inverse Hessians) from scratch at every iteration, we apply a simple
modification that combines the most recently observed information about the objective
function with the existing knowledge embedded in our current Hessian approximation.
The DFP updating formula is quite effective, but it was soon superseded by the BFGS
formula, which is presently considered to be the most effective of all quasi-Newton updating
formulae. BFGS updating can be derived by making a simple change in the argument
that led to (6.13). Instead of imposing conditions on the Hessian approximations Bk , we
impose similar conditions on their inverses Hk . The updated approximation Hk+1 must be
symmetric and positive definite, and must satisfy the secant equation (6.6), now written as
Hk+1 yk sk .
The condition of closeness to Hk is now specified by the following analogue of (6.9):
min H − Hk (6.16a)
H
subject to
H HT,
H yk sk .
(6.16b)
139
140
CHAPTER 6.
QUASI-NEWTON METHODS
The norm is again the weighted Frobenius norm described above, where the weight matrix
W is now any matrix satisfying W sk yk . (For concreteness, we assume again that W is
given by the average Hessian Ḡ k defined in (6.11).) The unique solution Hk+1 to (6.16) is
given by
(BFGS)
Hk+1 (I − ρk sk ykT )Hk (I − ρk yk skT ) + ρk sk skT ,
(6.17)
with ρk defined by (6.14).
Just one issue has to be resolved before we can define a complete BFGS algorithm: How
should we choose the initial approximation H0 ? Unfortunately, there is no magic formula
that works well in all cases. We can use specific information about the problem, for instance
by setting it to the inverse of an approximate Hessian calculated by finite differences at x0 .
Otherwise, we can simply set it to be the identity matrix, or a multiple of the identity matrix,
where the multiple is chosen to reflect the scaling of the variables.
Algorithm 6.1 (BFGS Method).
Given starting point x0 , convergence tolerance > 0,
inverse Hessian approximation H0 ;
k ← 0;
while ∇ f k > ;
Compute search direction
pk −Hk ∇ f k ;
(6.18)
Set xk+1 xk + αk pk where αk is computed from a line search
procedure to satisfy the Wolfe conditions (3.6);
Define sk xk+1 − xk and yk ∇ f k+1 − ∇ f k ;
Compute Hk+1 by means of (6.17);
k ← k + 1;
end (while)
Each iteration can be performed at a cost of O(n 2 ) arithmetic operations (plus the cost
of function and gradient evaluations); there are no O(n 3 ) operations such as linear system
solves or matrix–matrix operations. The algorithm is robust, and its rate of convergence is
superlinear, which is fast enough for most practical purposes. Even though Newton’s method
converges more rapidly (that is, quadratically), its cost per iteration usually is higher, because
of its need for second derivatives and solution of a linear system.
We can derive a version of the BFGS algorithm that works with the Hessian approximation Bk rather than Hk . The update formula for Bk is obtained by simply applying the
Sherman–Morrison–Woodbury formula (A.28) to (6.17) to obtain
(BFGS)
Bk+1 Bk −
Bk sk skT Bk
yk y T
+ Tk .
T
sk Bk sk
yk sk
(6.19)
6.1.
THE BFGS METHOD
A naive implementation of this variant is not efficient for unconstrained minimization,
because it requires the system Bk pk −∇ f k to be solved for the step pk , thereby increasing
the cost of the step computation to O(n 3 ). We discuss later, however, that less expensive
implementations of this variant are possible by updating Cholesky factors of Bk .
PROPERTIES OF THE BFGS METHOD
It is usually easy to observe the superlinear rate of convergence of the BFGS method
on practical problems. Below, we report the last few iterations of the steepest descent, BFGS,
and an inexact Newton method on Rosenbrock’s function (2.22). The table gives the value of
xk − x ∗ . The Wolfe conditions were imposed on the step length in all three methods. From
the starting point (−1.2, 1), the steepest descent method required 5264 iterations, whereas
BFGS and Newton took only 34 and 21 iterations, respectively to reduce the gradient norm
to 10−5 .
steepest
descent
1.827e-04
1.826e-04
1.824e-04
1.823e-04
BFGS
Newton
1.70e-03
1.17e-03
1.34e-04
1.01e-06
3.48e-02
1.44e-02
1.82e-04
1.17e-08
A few points in the derivation of the BFGS and DFP methods merit further discussion.
Note that the minimization problem (6.16) that gives rise to the BFGS update formula does
not explicitly require the updated Hessian approximation to be positive definite. It is easy to
show, however, that Hk+1 will be positive definite whenever Hk is positive definite, by using
the following argument. First, note from (6.8) that ykT sk is positive, so that the updating
formula (6.17), (6.14) is well-defined. For any nonzero vector z, we have
z T Hk+1 z w T Hk w + ρk (z T sk )2 ≥ 0,
where we have defined w z − ρk yk (skT z). The right hand side can be zero only if skT z 0,
but in this case w z 0, which implies that the first term is greater than zero. Therefore,
Hk+1 is positive definite.
To make quasi-Newton updating formulae invariant to transformations in the variables (such as scaling transformations), it is necessary for the objectives (6.9a) and (6.16a)
to be invariant under the same transformations. The choice of the weighting matrices W
used to define the norms in (6.9a) and (6.16a) ensures that this condition holds. Many other
choices of the weighting matrix W are possible, each one of them giving a different update
formula. However, despite intensive searches, no formula has been found that is significantly
more effective than BFGS.
141
142
CHAPTER 6.
QUASI-NEWTON METHODS
The BFGS method has many interesting properties when applied to quadratic functions. We discuss these properties later in the more general context of the Broyden family of
updating formulae, of which BFGS is a special case.
It is reasonable to ask whether there are situations in which the updating formula such
as (6.17) can produce bad results. If at some iteration the matrix Hk becomes a poor approximation to the true inverse Hessian, is there any hope of correcting it? For example, when
the inner product ykT sk is tiny (but positive), then it follows from (6.14), (6.17) that Hk+1
contains very large elements. Is this behavior reasonable? A related question concerns the
rounding errors that occur in finite-precision implementation of these methods. Can these
errors grow to the point of erasing all useful information in the quasi-Newton approximate
Hessian?
These questions have been studied analytically and experimentally, and it is now
known that the BFGS formula has very effective self-correcting properties. If the matrix Hk
incorrectly estimates the curvature in the objective function, and if this bad estimate slows
down the iteration, then the Hessian approximation will tend to correct itself within a few
steps. It is also known that the DFP method is less effective in correcting bad Hessian approximations; this property is believed to be the reason for its poorer practical performance. The
self-correcting properties of BFGS hold only when an adequate line search is performed. In
particular, the Wolfe line search conditions ensure that the gradients are sampled at points
that allow the model (6.1) to capture appropriate curvature information.
It is interesting to note that the DFP and BFGS updating formulae are duals of each
other, in the sense that one can be obtained from the other by the interchanges s ↔ y,
B ↔ H . This symmetry is not surprising, given the manner in which we derived these
methods above.
IMPLEMENTATION
A few details and enhancements need to be added to Algorithm 6.1 to produce an
efficient implementation. The line search, which should satisfy either the Wolfe conditions
(3.6) or the strong Wolfe conditions (3.7), should always try the step length αk 1 first,
because this step length will eventually always be accepted (under certain conditions), thereby
producing superlinear convergence of the overall algorithm. Computational observations
strongly suggest that it is more economical, in terms of function evaluations, to perform
a fairly inaccurate line search. The values c1 10−4 and c2 0.9 are commonly used in
(3.6).
As mentioned earlier, the initial matrix H0 often is set to some multiple β I of the
identity, but there is no good general strategy for choosing the multiple β. If β is too large,
so that the first step p0 −βg0 is too long, many function evaluations may be required to
find a suitable value for the step length α0 . Some software asks the user to prescribe a value
δ for the norm of the first step, and then set H0 δg0 −1 I to achieve this norm.
A heuristic that is often quite effective is to scale the starting matrix after the first
step has been computed but before the first BFGS update is performed. We change the
6.1.
THE BFGS METHOD
provisional value H0 I by setting
H0 ←
ykT sk
I,
ykT yk
(6.20)
before applying the update (6.14) , (6.17) to obtain H1 . This formula attempts to make the
size of H0 similar to that of ∇ 2 f (x0 )−1 , in the following sense. Assuming that the average
1/2
Hessian defined in (6.11) is positive definite, there exists a square root Ḡ k satisfying
1/2 1/2
1/2
Ḡ k Ḡ k Ḡ k (see Exercise 6.6). Therefore, by defining z k Ḡ k sk and using the relation
(6.12), we have
1/2
1/2
ykT sk
(Ḡ k sk )T Ḡ k sk
z kT z k
.
1/2
1/2
ykT yk
z kT Ḡ k z k
(Ḡ k sk )T Ḡ k Ḡ k sk
(6.21)
The reciprocal of (6.21) is an approximation to one of the eigenvalues of Ḡ k , which in turn
is close to an eigenvalue of ∇ 2 f (xk ). Hence, the quotient (6.21) itself approximates an
eigenvalue of ∇ 2 f (xk )−1 . Other scaling factors can be used in (6.20), but the one presented
here appears to be the most successful in practice.
In (6.19) we gave an update formula for a BFGS method that works with the Hessian approximation Bk instead of the the inverse Hessian approximation Hk . An efficient
implementation of this approach does not store Bk explicitly, but rather the Cholesky factorization L k Dk L kT of this matrix. A formula that updates the factors L k and Dk directly in
O(n 2 ) operations can be derived from (6.19). Since the linear system Bk pk −∇ f k also
can be solved in O(n 2 ) operations (by performing triangular substitutions with L k and L kT
and a diagonal substitution with Dk ), the total cost is quite similar to the variant described
in Algorithm 6.1. A potential advantage of this alternative strategy is that it gives us the
option of modifying diagonal elements in the Dk factor if they are not sufficiently large, to
prevent instability when we divide by these elements during the calculation of pk . However,
computational experience suggests no real advantages for this variant, and we prefer the
simpler strategy of Algorithm 6.1.
The performance of the BFGS method can degrade if the line search is not based
on the Wolfe conditions. For example, some software implements an Armijo backtracking
line search (see Section 3.1): The unit step length αk 1 is tried first and is successively
decreased until the sufficient decrease condition (3.6a) is satisfied. For this strategy, there is
no guarantee that the curvature condition ykT sk > 0 (6.7) will be satisfied by the chosen step,
since a step length greater than 1 may be required to satisfy this condition. To cope with this
shortcoming, some implementations simply skip the BFGS update by setting Hk+1 Hk
when ykT sk is negative or too close to zero. This approach is not recommended, because
the updates may be skipped much too often to allow Hk to capture important curvature
information for the objective function f . In Chapter 18 we discuss a damped BFGS update
that is a more effective strategy for coping with the case where the curvature condition (6.7)
is not satisfied.
143
144
CHAPTER 6.
6.2
QUASI-NEWTON METHODS
THE SR1 METHOD
In the BFGS and DFP updating formulae, the updated matrix Bk+1 (or Hk+1 ) differs from its
predecessor Bk (or Hk ) by a rank-2 matrix. In fact, as we now show, there is a simpler rank-1
update that maintains symmetry of the matrix and allows it to satisfy the secant equation.
Unlike the rank-two update formulae, this symmetric-rank-1, or SR1, update does not
guarantee that the updated matrix maintains positive definiteness. Good numerical results
have been obtained with algorithms based on SR1, so we derive it here and investigate its
properties.
The symmetric rank-1 update has the general form
Bk+1 Bk + σ vv T ,
where σ is either +1 or −1, and σ and v are chosen so that Bk+1 satisfies the secant equation
(6.6), that is, yk Bk+1 sk . By substituting into this equation, we obtain
yk Bk sk + σ v T sk v.
(6.22)
Since the term in brackets is a scalar, we deduce that v must be a multiple of yk − Bk sk , that
is, v δ(yk − Bk sk ) for some scalar δ. By substituting this form of v into (6.22), we obtain
(yk − Bk sk ) σ δ 2 skT (yk − Bk sk ) (yk − Bk sk ),
(6.23)
and it is clear that this equation is satisfied if (and only if) we choose the parameters δ and
σ to be
σ sign skT (yk − Bk sk ) ,
−1/2
δ ± skT (yk − Bk sk )
.
Hence, we have shown that the only symmetric rank-1 updating formula that satisfies the
secant equation is given by
(SR1)
Bk+1 Bk +
(yk − Bk sk )(yk − Bk sk )T
.
(yk − Bk sk )T sk
(6.24)
By applying the Sherman–Morrison formula (A.27), we obtain the corresponding update
formula for the inverse Hessian approximation Hk :
(SR1)
Hk+1 Hk +
(sk − Hk yk )(sk − Hk yk )T
.
(sk − Hk yk )T yk
(6.25)
This derivation is so simple that the SR1 formula has been rediscovered a number of times.
It is easy to see that even if Bk is positive definite, Bk+1 may not have the same property.
(The same is, of course, true of Hk .) This observation was considered a major drawback
6.2.
THE SR1 METHOD
in the early days of nonlinear optimization when only line search iterations were used.
However, with the advent of trust-region methods, the SR1 updating formula has proved to
be quite useful, and its ability to generate indefinite Hessian approximations can actually be
regarded as one of its chief advantages.
The main drawback of SR1 updating is that the denominator in (6.24) or (6.25) can
vanish. In fact, even when the objective function is a convex quadratic, there may be steps
on which there is no symmetric rank-1 update that satisfies the secant equation. It pays to
reexamine the derivation above in the light of this observation.
By reasoning in terms of Bk (similar arguments can be applied to Hk ), we see that
there are three cases:
1. If (yk − Bk sk )T sk 0, then the arguments above show that there is a unique
rank-one updating formula satisfying the secant equation (6.6), and that it is given
by (6.24).
2. If yk Bk sk , then the only updating formula satisfying the secant equation is simply
Bk+1 Bk .
3. If yk Bk sk and (yk − Bk sk )T sk 0, then (6.23) shows that there is no symmetric
rank-one updating formula satisfying the secant equation.
The last case clouds an otherwise simple and elegant derivation, and suggests that numerical
instabilities and even breakdown of the method can occur. It suggests that rank-one updating
does not provide enough freedom to develop a matrix with all the desired characteristics,
and that a rank-two correction is required. This reasoning leads us back to the BFGS method,
in which positive definiteness (and thus nonsingularity) of all Hessian approximations is
guaranteed.
Nevertheless, we are interested in the SR1 formula for the following reasons.
(i) A simple safeguard seems to adequately prevent the breakdown of the method and the
occurrence of numerical instabilities.
(ii) The matrices generated by the SR1 formula tend to be good approximations to the
true Hessian matrix—often better than the BFGS approximations.
(iii) In quasi-Newton methods for constrained problems, or in methods for partially
separable functions (see Chapters 18 and 7), it may not be possible to impose the
curvature condition ykT sk > 0, and thus BFGS updating is not recommended. Indeed,
in these two settings, indefinite Hessian approximations are desirable insofar as they
reflect indefiniteness in the true Hessian.
We now introduce a strategy to prevent the SR1 method from breaking down. It
has been observed in practice that SR1 performs well simply by skipping the update if the
denominator is small. More specifically, the update (6.24) is applied only if
T
s (yk − Bk sk ) ≥ r sk yk − Bk sk ,
k
(6.26)
145
146
CHAPTER 6.
QUASI-NEWTON METHODS
where r ∈ (0, 1) is a small number, say r 10−8 . If (6.26) does not hold, we set
Bk+1 Bk . Most implementations of the SR1 method use a skipping rule of this
kind.
Why do we advocate skipping of updates for the SR1 method, when in the previous
section we discouraged this strategy in the case of BFGS? The two cases are quite different.
The condition skT (yk − Bk sk ) ≈ 0 occurs infrequently, since it requires certain vectors to
be aligned in a specific way. When it does occur, skipping the update appears to have no
negative effects on the iteration. This is not surprising, since the skipping condition implies that skT Ḡsk ≈ skT Bk sk , where Ḡ is the average Hessian over the last step—meaning
that the curvature of Bk along sk is already correct. In contrast, the curvature condition
skT yk ≥ 0 required for BFGS updating may easily fail if the line search does not impose the Wolfe conditions (for example, if the step is not long enough), and therefore
skipping the BFGS update can occur often and can degrade the quality of the Hessian
approximation.
We now give a formal description of an SR1 method using a trust-region framework,
which we prefer over a line search framework because it can accommodate indefinite Hessian
approximations more easily.
Algorithm 6.2 (SR1 Trust-Region Method).
Given starting point x0 , initial Hessian approximation B0 ,
trust-region radius 0 , convergence tolerance > 0,
parameters η ∈ (0, 10−3 ) and r ∈ (0, 1);
k ← 0;
while ∇ f k > ;
Compute sk by solving the subproblem
1
min ∇ f kT s + s T Bk s
s
2
subject to s ≤ k ;
Compute
yk ∇ f (xk + sk ) − ∇ f k ,
ared f k − f (xk + sk )
(actual reduction)
1
T
T
(predicted reduction);
pred − ∇ f k sk + sk Bk sk
2
if ared/pred > η
xk+1 xk + sk ;
else
xk+1 xk ;
end (if)
(6.27)
6.2.
THE SR1 METHOD
if ared/pred > 0.75
if sk ≤ 0.8k
k+1 k ;
else
k+1 2k ;
end (if)
else if 0.1 ≤ ared/pred ≤ 0.75
k+1 k ;
else
k+1 0.5k ;
end (if)
if (6.26) holds
Use (6.24) to compute Bk+1 (even if xk+1 xk );
else
Bk+1 ← Bk ;
end (if)
k ← k + 1;
end (while)
This algorithm has the typical form of a trust region method (cf. Algorithm 4.1). For
concreteness, we have specified a particular strategy for updating the trust region radius,
but other heuristics can be used instead.
To obtain a fast rate of convergence, it is important for the matrix Bk to be updated
even along a failed direction sk . The fact that the step was poor indicates that Bk is an
inadequate approximation of the true Hessian in this direction. Unless the quality of the
approximation is improved, steps along similar directions could be generated on later
iterations, and repeated rejection of such steps could prevent superlinear convergence.
PROPERTIES OF SR1 UPDATING
One of the main advantages of SR1 updating is its ability to generate good Hessian
approximations. We demonstrate this property by first examining a quadratic function. For
functions of this type, the choice of step length does not affect the update, so to examine the
effect of the updates, we can assume for simplicity a uniform step length of 1, that is,
pk −Hk ∇ f k ,
xk+1 xk + pk .
(6.28)
It follows that pk sk .
Theorem 6.1.
Suppose that f : IRn → IR is the strongly convex quadratic function f (x) b T x +
1 T
x Ax, where A is symmetric positive definite. Then for any starting point x0 and any
2
147
148
CHAPTER 6.
QUASI-NEWTON METHODS
symmetric starting matrix H0 , the iterates {xk } generated by the SR1 method (6.25), (6.28)
converge to the minimizer in at most n steps, provided that (sk − Hk yk )T yk 0 for all k.
Moreover, if n steps are performed, and if the search directions pi are linearly independent, then
Hn A−1 .
PROOF. Because of our assumption (sk − Hk yk )T yk 0, the SR1 update is always welldefined. We start by showing inductively that
Hk y j s j ,
j 0, 1, . . . , k − 1.
(6.29)
In other words, we claim that the secant equation is satisfied not only along the most recent
search direction, but along all previous directions.
By definition, the SR1 update satisfies the secant equation, so we have H1 y0 s0 . Let
us now assume that (6.29) holds for some value k > 1 and show that it holds also for k + 1.
From this assumption, we have from (6.29) that
(sk − Hk yk )T y j skT y j − ykT (Hk y j ) skT y j − ykT s j 0,
all j < k,
(6.30)
where the last equality follows because yi Asi for the quadratic function we are considering
here. By using (6.30) and the induction hypothesis (6.29) in (6.25), we have
Hk+1 y j Hk y j s j ,
for all j < k.
Since Hk+1 yk sk by the secant equation, we have shown that (6.29) holds when k is
replaced by k + 1. By induction, then, this relation holds for all k.
If the algorithm performs n steps and if these steps {s j } are linearly independent, we
have
s j Hn y j Hn As j ,
j 0, 1, . . . , n − 1.
It follows that Hn A I , that is, Hn A−1 . Therefore, the step taken at xn is the Newton
step, and so the next iterate xn+1 will be the solution, and the algorithm terminates.
Consider now the case in which the steps become linearly dependent. Suppose that sk
is a linear combination of the previous steps, that is,
sk ξ0 s0 + · · · + ξk−1 sk−1 ,
for some scalars ξi . From (6.31) and (6.29) we have that
Hk yk Hk Ask
ξ0 Hk As0 + · · · + ξk−1 Hk Ask−1
(6.31)
6.3.
THE BROYDEN CLASS
ξ0 Hk y0 + · · · + ξk−1 Hk yk−1
ξ0 s0 + · · · + ξk−1 sk−1
sk .
Since yk ∇ f k+1 − ∇ f k and since sk pk −Hk ∇ f k from (6.28), we have that
Hk (∇ f k+1 − ∇ f k ) −Hk ∇ f k ,
which, by the nonsingularity of Hk , implies that ∇ f k+1 0. Therefore, xk+1 is the solution
point.
The relation (6.29) shows that when f is quadratic, the secant equation is satisfied
along all previous search directions, regardless of how the line search is performed. A result
like this can be established for BFGS updating only under the restrictive assumption that
the line search is exact, as we show in the next section.
For general nonlinear functions, the SR1 update continues to generate good Hessian
approximations under certain conditions.
Theorem 6.2.
Suppose that f is twice continuously differentiable, and that its Hessian is bounded and
Lipschitz continuous in a neighborhood of a point x ∗ . Let {xk } be any sequence of iterates such
that xk → x ∗ for some x ∗ ∈ IRn . Suppose in addition that the inequality (6.26) holds for all k,
for some r ∈ (0, 1), and that the steps sk are uniformly linearly independent. Then the matrices
Bk generated by the SR1 updating formula satisfy
lim Bk − ∇ 2 f (x ∗ ) 0.
k→∞
The term “uniformly linearly independent steps” means, roughly speaking, that the
steps do not tend to fall in a subspace of dimension less than n. This assumption is usually,
but not always, satisfied in practice (see the Notes and References at the end of this chapter).
6.3
THE BROYDEN CLASS
So far, we have described the BFGS, DFP, and SR1 quasi-Newton updating formulae, but
there are many others. Of particular interest is the Broyden class, a family of updates specified
by the following general formula:
Bk+1 Bk −
Bk sk skT Bk
yk y T
+ T k + φk (skT Bk sk )vk vkT ,
T
sk Bk sk
yk sk
(6.32)
149
150
CHAPTER 6.
QUASI-NEWTON METHODS
where φk is a scalar parameter and
vk yk
Bk sk
.
− T
ykT sk
sk Bk sk
(6.33)
The BFGS and DFP methods are members of the Broyden class—we recover BFGS by setting
φk 0 and DFP by setting φk 1 in (6.32). We can therefore rewrite (6.32) as a “linear
combination” of these two methods, that is,
BFGS
DFP
+ φk Bk+1
.
Bk+1 (1 − φk )Bk+1
This relationship indicates that all members of the Broyden class satisfy the secant equation
(6.6), since the BGFS and DFP matrices themselves satisfy this equation. Also, since BFGS and
DFP updating preserve positive definiteness of the Hessian approximations when skT yk > 0,
this relation implies that the same property will hold for the Broyden family if 0 ≤ φk ≤ 1.
Much attention has been given to the so-called restricted Broyden class, which is
obtained by restricting φk to the interval [0, 1]. It enjoys the following property when
applied to quadratic functions. Since the analysis is independent of the step length, we
assume for simplicity that each iteration has the form
pk −Bk−1 ∇ f k ,
xk+1 xk + pk .
(6.34)
Theorem 6.3.
Suppose that f : IRn → IR is the strongly convex quadratic function f (x) b T x +
1 T
x Ax, where A is symmetric and positive definite. Let x0 be any starting point for the iteration
2
(6.34) and B0 be any symmetric positive definite starting matrix, and suppose that the matrices
Bk are updated by the Broyden formula (6.32) with φk ∈ [0, 1]. Define λk1 ≤ λk2 ≤ · · · ≤ λkn
to be the eigenvalues of the matrix
1
1
A 2 Bk−1 A 2 .
(6.35)
Then for all k, we have
min{λik , 1} ≤ λik+1 ≤ max{λik , 1},
i 1, 2, . . . , n.
(6.36)
Moreover, the property (6.36) does not hold if the Broyden parameter φk is chosen outside the
interval [0, 1].
Let us discuss the significance of this result. If the eigenvalues λik of the matrix (6.35)
are all 1, then the quasi-Newton approximation Bk is identical to the Hessian A of the
quadratic objective function. This situation is the ideal one, so we should be hoping for
these eigenvalues to be as close to 1 as possible. In fact, relation (6.36) tells us that the
6.3.
THE BROYDEN CLASS
eigenvalues {λik } converge monotonically (but not strictly monotonically) to 1. Suppose, for
example, that at iteration k the smallest eigenvalue is λk1 0.7. Then (6.36) tells us that
∈ [0.7, 1]. We cannot be sure that this eigenvalue has actually
at the next iteration λk+1
1
moved closer to 1, but it is reasonable to expect that it has. In contrast, the first eigenvalue
can become smaller than 0.7 if we allow φk to be outside [0, 1]. Significantly, the result of
Theorem 6.3 holds even if the line searches are not exact.
Although Theorem 6.3 seems to suggest that the best update formulas belong to the
restricted Broyden class, the situation is not at all clear. Some analysis and computational
testing suggest that algorithms that allow φk to be negative (in a strictly controlled manner)
may in fact be superior to the BFGS method. The SR1 formula is a case in point: It is a
member of the Broyden class, obtained by setting
φk skT yk
,
skT yk − skT Bk sk
but it does not belong to the restricted Broyden class, because this value of φk may fall
outside the interval [0, 1].
In the remaining discussion of this section, we determine more precisely the range of
values of φk that preserve positive definiteness.
The last term in (6.32) is a rank-one correction, which by the interlacing eigenvalue
theorem (Theorem A.1) increases the eigenvalues of the matrix when φk is positive. Therefore
Bk+1 is positive definite for all φk ≥ 0. On the other hand, by Theorem A.1 the last term in
(6.32) decreases the eigenvalues of the matrix when φk is negative. As we decrease φk , this
matrix eventually becomes singular and then indefinite. A little computation shows that
Bk+1 is singular when φk has the value
1
,
1 − µk
(6.37)
(ykT Bk−1 yk )(skT Bk sk )
.
(ykT sk )2
(6.38)
φkc where
µk By applying the Cauchy–Schwarz inequality (A.5) to (6.38), we see that µk ≥ 1 and therefore
φkc ≤ 0. Hence, if the initial Hessian approximation B0 is symmetric and positive definite,
and if skT yk > 0 and φk > φkc for each k, then all the matrices Bk generated by Broyden’s
formula (6.32) remain symmetric and positive definite.
When the line search is exact, all methods in the Broyden class with φk ≥ φkc generate
the same sequence of iterates. This result applies to general nonlinear functions and is
based on the observation that when all the line searches are exact, the directions generated
by Broyden-class methods differ only in their lengths. The line searches identify the same
151
152
CHAPTER 6.
QUASI-NEWTON METHODS
minima along the chosen search direction, though the values of the step lengths may differ
because of the different scaling.
The Broyden class has several remarkable properties when applied with exact line
searches to quadratic functions. We state some of these properties in the next theorem,
whose proof is omitted.
Theorem 6.4.
Suppose that a method in the Broyden class is applied to the strongly convex quadratic
function f (x) b T x + 12 x T Ax, where x0 is the starting point and B0 is any symmetric positive
definite matrix. Assume that αk is the exact step length and that φk ≥ φkc for all k, where φkc is
defined by (6.37). Then the following statements are true.
(i) The iterates are independent of φk and converge to the solution in at most n iterations.
(ii) The secant equation is satisfied for all previous search directions, that is,
Bk s j y j ,
j k − 1, k − 2, . . . , 1.
(iii) If the starting matrix is B0 I , then the iterates are identical to those generated by
the conjugate gradient method (see Chapter 5). In particular, the search directions are
conjugate, that is,
siT As j 0,
for i j.
(iv) If n iterations are performed, we have Bn A.
Note that parts (i), (ii), and (iv) of this result echo the statement and proof of Theorem 6.1,
where similar results were derived for the SR1 update formula.
We can generalize Theorem 6.4 slightly: It continues to hold if the Hessian approximations remain nonsingular but not necessarily positive definite. (Hence, we could allow
φk to be smaller than φkc , provided that the chosen value did not produce a singular updated
matrix.) We can also generalize point (iii) as follows. If the starting matrix B0 is not the
identity matrix, then the Broyden-class method is identical to the preconditioned conjugate
gradient method that uses B0 as preconditioner.
We conclude by commenting that results like Theorem 6.4 would appear to be of
mainly theoretical interest, since the inexact line searches used in practical implementations
of Broyden-class methods (and all other quasi-Newton methods) cause their performance
to differ markedly. Nevertheless, it is worth noting that this type of analysis guided much of
the development of quasi-Newton methods.
6.4.
6.4
CONVERGENCE ANALYSIS
CONVERGENCE ANALYSIS
In this section we present global and local convergence results for practical implementations
of the BFGS and SR1 methods. We give more details for BFGS because its analysis is more
general and illuminating than that of SR1. The fact that the Hessian approximations evolve
by means of updating formulas makes the analysis of quasi-Newton methods much more
complex than that of steepest descent and Newton’s method.
Although the BFGS and SR1 methods are known to be remarkably robust in practice,
we will not be able to establish truly global convergence results for general nonlinear objective
functions. That is, we cannot prove that the iterates of these quasi-Newton methods approach
a stationary point of the problem from any starting point and any (suitable) initial Hessian
approximation. In fact, it is not yet known if the algorithms enjoy such properties. In our
analysis we will either assume that the objective function is convex or that the iterates satisfy
certain properties. On the other hand, there are well known local, superlinear convergence
results that are true under reasonable assumptions.
Throughout this section we use · to denote the Euclidean vector or matrix norm,
and denote the Hessian matrix ∇ 2 f (x) by G(x).
GLOBAL CONVERGENCE OF THE BFGS METHOD
We study the global convergence of the BFGS method, with a practical line search,
when applied to a smooth convex function from an arbitrary starting point x0 and from
any initial Hessian approximation B0 that is symmetric and positive definite. We state our
precise assumptions about the objective function formally, as follows.
Assumption 6.1.
(i) The objective function f is twice continuously differentiable.
(ii) The level set L {x ∈ IRn | f (x) ≤ f (x0 )} is convex, and there exist positive constants
m and M such that
mz2 ≤ z T G(x)z ≤ Mz2
(6.39)
for all z ∈ IRn and x ∈ L.
Part (ii) of this assumption implies that G(x) is positive definite on L and that f has a
unique minimizer x ∗ in L.
By using (6.12) and (6.39) we obtain
ykT sk
s T Ḡ k sk
kT
≥ m,
T
sk sk
sk sk
(6.40)
153
154
CHAPTER 6.
QUASI-NEWTON METHODS
where Ḡ k is the average Hessian defined in (6.11). Assumption 6.1 implies that Ḡ k is
positive definite, so its square root is well-defined. Therefore, as in (6.21), we have by
1/2
defining z k Ḡ k sk that
ykT yk
skT Ḡ 2k sk
z kT Ḡ k z k
≤ M.
ykT sk
z kT z k
skT Ḡ k sk
(6.41)
We are now ready to present the global convergence result for the BFGS method. It
does not seem to be possible to establish a bound on the condition number of the Hessian
approximations Bk , as is done in Section 3.2. Instead, we will introduce two new tools
in the analysis, the trace and determinant, to estimate the size of the largest and smallest
eigenvalues of the Hessian approximations. The trace of a matrix (denoted by trace(·)) is
the sum of its eigenvalues, while the determinant (denoted by det(·)) is the product of the
eigenvalues; see the Appendix for a brief discussion of their properties.
Theorem 6.5.
Let B0 be any symmetric positive definite initial matrix, and let x0 be a starting point
for which Assumption 6.1 is satisfied. Then the sequence {xk } generated by Algorithm 6.1 (with
0) converges to the minimizer x ∗ of f .
PROOF. We start by defining
mk ykT sk
,
skT sk
Mk ykT yk
,
ykT sk
(6.42)
and note from (6.40) and (6.41) that
m k ≥ m,
Mk ≤ M.
(6.43)
By computing the trace of the BFGS update (6.19), we obtain that
trace(Bk+1 ) trace(Bk ) −
Bk sk 2
yk 2
+ T
T
sk Bk sk
yk sk
(6.44)
(see Exercise 6.11). We can also show (Exercise 6.10) that
det(Bk+1 ) det(Bk )
ykT sk
.
skT Bk sk
(6.45)
We now define
cos θk skT Bk sk
,
sk Bk sk qk skT Bk sk
,
skT sk
(6.46)
6.4.
CONVERGENCE ANALYSIS
so that θk is the angle between sk and Bk sk . We then obtain that
Bk sk 2
Bk sk 2 sk 2 skT Bk sk
qk
.
T
T
2
2
sk cos2 θk
sk Bk sk
(sk Bk sk )
(6.47)
In addition, we have from (6.42) that
det(Bk+1 ) det(Bk )
ykT sk skT sk
mk
det(Bk ) .
T
T
qk
sk sk sk Bk sk
(6.48)
We now combine the trace and determinant by introducing the following function of
a positive definite matrix B:
ψ(B) trace(B) − ln(det(B)),
(6.49)
where ln(·) denotes the natural logarithm. It is not difficult to show that ψ(B) > 0; see
Exercise 6.9. By using (6.42) and (6.44)–(6.49), we have that
qk
− ln(det(Bk )) − ln m k + ln qk
cos2 θk
ψ(Bk ) + (Mk − ln m k − 1)
qk
qk
+ ln cos2 θk .
+ 1−
+ ln
cos2 θk
cos2 θk
ψ(Bk+1 ) trace(Bk ) + Mk −
(6.50)
Now, since the function h(t) 1 − t + ln t is nonpositive for all t > 0 (see Exercise 6.8),
the term inside the square brackets is nonpositive, and thus from (6.43) and (6.50) we have
0 < ψ(Bk+1 ) ≤ ψ(B0 ) + c(k + 1) +
k
ln cos2 θ j ,
(6.51)
j0
where we can assume the constant c M −ln m −1 to be positive, without loss of generality.
We now relate these expressions to the results given in Section 3.2. Note from the form
sk −αk Bk−1 ∇ f k of the quasi-Newton iteration that cos θk defined by (6.46) is the angle
between the steepest descent direction and the search direction, which plays a crucial role in
the global convergence theory of Chapter 3. From (3.22), (3.23) we know that the sequence
∇ f k generated by the line search algorithm is bounded away from zero only if cos θ j → 0.
Let us then proceed by contradiction and assume that cos θ j → 0. Then there exists
k1 > 0 such that for all j > k1 , we have
ln cos2 θ j < −2c,
155
156
CHAPTER 6.
QUASI-NEWTON METHODS
where c is the constant defined above. Using this inequality in (6.51) we find the following
relations to be true for all k > k1 :
0 < ψ(B0 ) + c(k + 1) +
k1
ln cos2 θ j +
j0
ψ(B0 ) +
k1
k
(−2c)
jk1 +1
ln cos2 θ j + 2ck1 + c − ck.
j0
However, the right-hand-side is negative for large k, giving a contradiction. Therefore, there
exists a subsequence of indices { jk }k1,2,... such that cos θ jk ≥ δ > 0. By Zoutendijk’s result
(3.14) this limit implies that lim inf ∇ f k → 0. Since the problem is strongly convex, the
latter limit is enough to prove that xk → x ∗ .
Theorem 6.5 has been generalized to the entire restricted Broyden class, except for
the DFP method. In other words, Theorem 6.5 can be shown to hold for all φk ∈ [0, 1)
in (6.32), but the argument seems to break down as φk approaches 1 because some of the
self-correcting properties of the update are weakened considerably.
An extension of the analysis just given shows that the rate of convergence of the iterates
is linear. In particular, we can show that the sequence xk − x ∗ converges to zero rapidly
enough that
∞
xk − x ∗ < ∞.
(6.52)
k1
We will not prove this claim, but rather establish that if (6.52) holds, then the rate of
convergence is actually superlinear.
SUPERLINEAR CONVERGENCE OF THE BFGS METHOD
The analysis of this section makes use of the Dennis and Moré characterization (3.36)
of superlinear convergence. It applies to general nonlinear—not just convex—objective
functions. For the results that follow we need to make an additional assumption.
Assumption 6.2.
The Hessian matrix G is Lipschitz continuous at x ∗ , that is,
G(x) − G(x ∗ ) ≤ Lx − x ∗ ,
for all x near x ∗ , where L is a positive constant.
6.4.
CONVERGENCE ANALYSIS
We start by introducing the quantities
ỹk G −1/2
yk ,
∗
s̃k G 1/2
∗ sk ,
B̃k G −1/2
Bk G −1/2
,
∗
∗
where G ∗ G(x ∗ ) and x ∗ is a minimizer of f . Similarly to (6.46), we define
cos θ̃k s̃kT B̃k s̃k
s̃k B̃k s̃k ,
q̃k s̃kT B̃k s̃k
,
s̃k 2
while we echo (6.42) and (6.43) in defining
M̃k ỹk 2
,
ỹkT s̃k
m̃ k ỹkT s̃k
.
s̃kT s̃k
−1/2
By pre- and postmultiplying the BFGS update formula (6.19) by G ∗
terms appropriately, we obtain
B̃k+1 B̃k −
B̃k s̃k s̃kT B̃k
s̃kT B̃k s̃k
+
and grouping
ỹk ỹkT
.
ỹkT s̃k
Since this expression has precisely the same form as the BFGS formula (6.19), it follows
from the argument leading to (6.50) that
ψ( B̃k+1 ) ψ( B̃k ) + ( M̃k − ln m̃ k − 1)
q̃k
q̃k
+ 1−
+ ln
cos2 θ̃k
cos2 θ̃k
2
+ ln cos θ̃k .
Recalling (6.12), we have that
yk − G ∗ sk (Ḡ k − G ∗ )sk ,
and thus
(Ḡ k − G ∗ )G −1/2
s̃k .
ỹk − s̃k G −1/2
∗
∗
By Assumption 6.2, and recalling the definition (6.11), we have
2 s˜k Ḡ k − G ∗ ≤ G −1/2
2 s̃k L k ,
ỹk − s̃k ≤ G −1/2
∗
∗
where
k
is defined by
k
max{xk+1 − x ∗ , xk − x ∗ }.
(6.53)
157
158
CHAPTER 6.
QUASI-NEWTON METHODS
We have thus shown that
ỹk − s̃k ≤ c̄ k ,
s̃k (6.54)
for some positive constant c̄. This inequality and (6.52) play an important role in superlinear
convergence, as we now show.
Theorem 6.6.
Suppose that f is twice continuously differentiable and that the iterates generated by the
BFGS algorithm converge to a minimizer x ∗ at which Assumption 6.2 holds. Suppose also that
(6.52) holds. Then xk converges to x ∗ at a superlinear rate.
PROOF. From (6.54), we have from the triangle inequality (A.4a) that
ỹk − s̃k ≤ c̄ k s̃k ,
s̃k − y˜k ≤ c̄ k s˜k ,
so that
(1 − c̄ k )s̃k ≤ ỹk ≤ (1 + c̄ k )s̃k .
(6.55)
By squaring (6.54) and using (6.55), we obtain
(1 − c̄ k )2 s̃k 2 − 2 ỹkT s̃k + s˜k 2 ≤ ỹk 2 − 2 ỹkT s̃k + s˜k 2 ≤ c̄2 k2 s̃k 2 ,
and therefore
2 ỹkT s̃k ≥ (1 − 2c̄
k
+ c̄2
+ 1 − c̄2 k2 )s̃k 2 2(1 − c̄ k )s˜k 2 .
2
k
It follows from the definition of m̃ k that
m̃ k ỹkT s̃k
≥ 1 − c̄ k .
s̃k 2
(6.56)
By combining (6.55) and (6.56), we obtain also that
M̃k y˜k 2
1 + c̄ k
≤
.
1 − c̄ k
ỹkT s̃k
(6.57)
Since xk → x ∗ , we have that k → 0, and thus by (6.57) there exists a positive constant
c > c̄ such that the following inequalities hold for all sufficiently large k:
M̃k ≤ 1 +
2c̄
1 − c̄
k
k
≤ 1 + c k.
(6.58)
6.4.
CONVERGENCE ANALYSIS
We again make use of the nonpositiveness of the function h(t) 1 − t + ln t. Therefore,
we have
−x
− ln(1 − x) h
1−x
Now, for k large enough we can assume that c̄
ln(1 − c̄ k ) ≥
k
1
1−x
≤ 0.
< 12 , and therefore
−c̄ k
≥ −2c̄ k .
1 − c̄ k
This relation and (6.56) imply that for sufficiently large k, we have
ln m̃ k ≥ ln(1 − c̄ k ) ≥ −2c̄
k
> −2c k .
(6.59)
We can now deduce from (6.53), (6.58), and (6.59) that
0 < ψ( B̃k+1 ) ≤ ψ( B̃k ) + 3c
k
˜
+ ln cos θk + 1 −
2
q˜k
cos2 θ˜k
+ ln
q˜k
.
cos2 θ˜k
(6.60)
By summing this expression and making use of (6.52) we have that
∞
j0
ln
1
cos2 θ˜j
− 1−
q˜j
cos2 θ˜j
+ ln
q˜j
≤ ψ( B̃0 ) + 3c
cos2 θ˜j
∞
j
< +∞.
j0
Since the term in the square brackets is nonpositive, and since ln 1/ cos2 θ̃ j ≥ 0 for all j,
we obtain the two limits
lim ln
j→∞
1
cos2 θ˜j
0,
lim
j→∞
1−
q˜j
cos2 θ˜j
+ ln
q˜j
cos2 θ˜j
0,
which imply that
lim cos θ˜j 1,
j→∞
lim q˜j 1.
j→∞
(6.61)
The essence of the result has now been proven; we need only to interpret these limits
in terms of the Dennis–Moré characterization of superlinear convergence.
159
160
CHAPTER 6.
QUASI-NEWTON METHODS
Recalling (6.47), we have
−1/2
G ∗
(Bk − G ∗ )sk 2
1/2
G ∗ sk 2
( B̃k − I )s̃k 2
s̃k 2
B̃k s̃k 2 − 2s̃kT B̃k s̃k + s̃kT s̃k
s̃kT s̃k
q̃k2
cos θ̃k2
− 2q̃k + 1.
Since by (6.61) the right-hand-side converges to 0, we conclude that
(Bk − G ∗ )sk 0.
k→∞
sk lim
The limit (3.36) and Theorem 3.6 imply that the unit step length αk 1 will satisfy the Wolfe
conditions near the solution, and hence that the rate of convergence is superlinear.
CONVERGENCE ANALYSIS OF THE SR1 METHOD
The convergence properties of the SR1 method are not as well understood as those of
the BFGS method. No global results like Theorem 6.5 or local superlinear results like Theorem 6.6 have been established, except the results for quadratic functions discussed earlier.
There is, however, an interesting result for the trust-region SR1 algorithm, Algorithm 6.2.
It states that when the objective function has a unique stationary point and the condition
(6.26) holds at every step (so that the SR1 update is never skipped) and the Hessian approximations Bk are bounded above, then the iterates converge to x ∗ at an (n + 1)-step
superlinear rate. The result does not require exact solution of the trust-region subproblem
(6.27).
We state the result formally as follows.
Theorem 6.7.
Suppose that the iterates xk are generated by Algorithm 6.2. Suppose also that the following
conditions hold:
(c1) The sequence of iterates does not terminate, but remains in a closed, bounded, convex set
D, on which the function f is twice continuously differentiable, and in which f has a
unique stationary point x ∗ ;
(c2) the Hessian ∇ 2 f (x ∗ ) is positive definite, and ∇ 2 f (x) is Lipschitz continuous in a
neighborhood of x ∗ ;
(c3) the sequence of matrices {Bk } is bounded in norm;
(c4) condition (6.26) holds at every iteration, where r is some constant in (0, 1).
6.4.
CONVERGENCE ANALYSIS
Then limk→∞ xk x ∗ , and we have that
xk+n+1 − x ∗ 0.
k→∞
xk − x ∗ lim
Note that the BFGS method does not require the boundedness assumption (c3) to
hold. As we have mentioned already, the SR1 update does not necessarily maintain positive
definiteness of the Hessian approximations Bk . In practice, Bk may be indefinite at any
iteration, which means that the trust region bound may continue to be active for arbitrarily
large k. Interestingly, however, it can be shown that the SR1 Hessian approximations tend
to be positive definite most of the time. The precise result is that
number of indices j 1, 2, . . . , k for which B j is positive semidefinite
1,
k→∞
k
lim
under the assumptions of Theorem 6.7. This result holds regardless of whether the initial
Hessian approximation is positive definite or not.
NOTES AND REFERENCES
For a comprehensive treatment of quasi-Newton methods see Dennis and Schnabel [92], Dennis and Moré [91], and Fletcher [101]. A formula for updating the Cholesky
factors of the BFGS matrices is given in Dennis and Schnabel [92].
Several safeguards and modifications of the SR1 method have been proposed, but
the condition (6.26) is favored in the light of the analysis of Conn, Gould, and Toint [71].
Computational experiments by Conn, Gould, and Toint [70, 73] and Khalfan, Byrd, and
Schnabel [181], using both line search and trust-region approaches, indicate that the SR1
method appears to be competitive with the BFGS method. The proof of Theorem 6.7 is
given in Byrd, Khalfan, and Schnabel [51].
A study of the convergence of BFGS matrices for nonlinear problems can be found in
Ge and Powell [119] and Boggs and Tolle [32]; however, the results are not as satisfactory as
for SR1 updating.
The global convergence of the BFGS method was established by Powell [246]. This
result was extended to the restricted Broyden class, except for DFP, by Byrd, Nocedal, and
Yuan [53]. For a discussion of the self-correcting properties of quasi-Newton methods
see Nocedal [229]. Most of the early analysis of quasi-Newton methods was based on the
bounded deterioration principle. This is a tool for the local analysis that quantifies the worstcase behavior of quasi-Newton updating. Assuming that the starting point is sufficiently
close to the solution x ∗ and that the initial Hessian approximation is sufficiently close to
∇ 2 f (x ∗ ), one can use the bounded deterioration bounds to prove that the iteration cannot
stray away from the solution. This property can then be used to show that the quality of the
quasi-Newton approximations is good enough to yield superlinear convergence. For details,
see Dennis and Moré [91] or Dennis and Schnabel [92].
161
162
CHAPTER 6.
QUASI-NEWTON METHODS
✐
EXERCISES
✐
6.1
(a) Show that if f is strongly convex, then (6.7) holds for any vectors xk and xk+1 .
(b) Give an example of a function of one variable satisfying g(0) −1 and g(1) − 14
and show that (6.7) does not hold in this case.
✐
6.2 Show that the second strong Wolfe condition (3.7b) implies the curvature
condition (6.7).
✐
6.3 Verify that (6.19) and (6.17) are inverses of each other.
✐
6.4 Use the Sherman–Morrison formula (A.27) to show that (6.24) is the inverse of
(6.25).
✐
6.5 Prove the statements (ii) and (iii) given in the paragraph following (6.25).
✐
6.6 The square root of a matrix A is a matrix A1/2 such that A1/2 A1/2 A. Show
that any symmetric positive definite matrix A has a square root, and that this square root
is itself symmetric and positive definite. (Hint: Use the factorization A U DU T (A.16),
where U is orthogonal and D is diagonal with positive diagonal elements.)
✐
6.7 Use the Cauchy-Schwarz inequality (A.5) to verify that µk ≥ 1, where µk is
defined by (6.38).
✐
6.8 Define h(t) 1 − t + ln t, and note that h (t) −1 + 1/t, h (t) −1/t 2 < 0,
h(1) 0, and h (1) 0. Show that h(t) ≤ 0 for all t > 0.
✐
6.9 Denote the eigenvalues of the positive definite matrix B by λ1 , λ2 , . . . , λn , where
0 < λ1 ≤ λ2 ≤ · · · ≤ λn . Show that the ψ function defined in (6.49) can be written as
ψ(B) n
(λi − ln λi ).
i1
Use this form to show that ψ(B) > 0.
✐
6.10 The object of this exercise is to prove (6.45).
(a) Show that det(I + x y T ) 1 + y T x, where x and y are n-vectors. Hint: Assuming that
x 0, we can find vectors w1 , w2 , . . . , wn−1 such that the matrix Q defined by
Q [x, w1 , w2 , . . . , wn−1 ]
6.4.
CONVERGENCE ANALYSIS
is nonsingular and x Qe1 , where e1 (1, 0, 0, . . . , 0)T . If we define
y T Q (z 1 , z 2 , . . . , z n ),
then
z 1 y T Qe1 y T Q(Q −1 x) y T x,
and
det(I + x y T ) det(Q −1 (I + x y T )Q) det(I + e1 y T Q).
(b) Use a similar technique to prove that
det(I + x y T + uv T ) (1 + y T x)(1 + v T u) − (x T v)(y T u).
(c) Use this relation to establish (6.45).
✐
6.11 Use the properties of the trace of a symmetric matrix and the formula (6.19) to
prove (6.44).
✐ 6.12 Show that if f satisfies Assumption 6.1 and if the sequence of gradients satisfies
lim inf ∇ f k 0, then the whole sequence of iterates x converges to the solution x ∗ .
163
This is pag
Printer: O
CHAPTER
7
Large-Scale
Unconstrained
Optimization
Many applications give rise to unconstrained optimization problems with thousands or
millions of variables. Problems of this size can be solved efficiently only if the storage
and computational costs of the optimization algorithm can be kept at a tolerable level. A
diverse collection of large-scale optimization methods has been developed to achieve this
goal, each being particularly effective for certain problem types. Some of these methods
are straightforward adaptations of the methods described in Chapters 3, 4, and 6. Other
approaches are modifications of these basic methods that allow approximate steps to be
calculated at lower cost in computation and storage. One set of approaches that we have
already discussed—the nonlinear conjugate gradient methods of Section 5.2—can be applied
7.1.
INEXACT NEWTON METHODS
to large problems without modification, because of its minimal storage demands and its
reliance on only first-order derivative information.
The line search and trust-region Newton algorithms of Chapters 3 and 4 require matrix
factorizations of the Hessian matrices ∇ 2 f k . In the large-scale case, these factorizations can
be carried out using sparse elimination techniques. Such algorithms have received much
attention, and high quality software implementations are available. If the computational
cost and memory requirements of these sparse factorization methods are affordable for a
given application, and if the Hessian matrix can be formed explicitly, Newton methods
based on sparse factorizations constitute an effective approach for solving such problems.
Often, however, the cost of factoring the Hessian is prohibitive, and it is preferable
to compute approximations to the Newton step using iterative linear algebra techniques.
Section 7.1 discusses inexact Newton methods that use these techniques, in both line search
and trust-region frameworks. The resulting algorithms have attractive global convergence
properties and may be superlinearly convergent for suitable choices of parameters. They find
effective search directions when the Hessian ∇ 2 f k is indefinite, and may even be implemented
in a “Hessian-free” manner, without explicit calculation or storage of the Hessian.
The Hessian approximations generated by the quasi-Newton approaches of Chapter 6
are usually dense, even when the true Hessian is sparse, and the cost of storing and working
with these approximations can be excessive for large n. Section 7.2 discusses limited-memory
variants of the quasi-Newton approach, which use Hessian approximations that can be
stored compactly by using just a few vectors of length n. These methods are fairly robust,
inexpensive, and easy to implement, but they do not converge rapidly. Another approach,
discussed briefly in Section 7.3, is to define quasi-Newton approximate Hessians Bk that
preserve sparsity, for example by mimicking the sparsity pattern of the Hessian.
In Section 7.4, we note that objective functions in large problems often possess a
structural property known as partial separability, which means they can be decomposed
into a sum of simpler functions, each of which depends on only a small subspace of IRn .
Effective Newton and quasi-Newton methods that exploit this property have been developed.
Such methods usually converge rapidly and are robust, but they require detailed information
about the objective function, which can be difficult to obtain in some applications.
We conclude the chapter with a discussion of software for large-scale unconstrained
optimization problems.
7.1
INEXACT NEWTON METHODS
Recall from (2.15) that the basic Newton step pkN is obtained by solving the symmetric n × n
linear system
∇ 2 f k pkN −∇ f k .
(7.1)
In this section, we describe techniques for obtaining approximations to pkN that are
165
166
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
inexpensive to calculate but are good search directions or steps. These approaches are
based on solving (7.1) by using the conjugate gradient (CG) method (see Chapter 5) or
the Lanczos method, with modifications to handle negative curvature in the Hessian ∇ 2 f k .
Both line search and trust-region approaches are described here. We refer to this family of
methods by the general name inexact Newton methods.
The use of iterative methods for (7.1) spares us from concerns about the expense
of a direct factorization of the Hessian ∇ 2 f k and the fill-in that may occur during this
process. Further, we can customize the solution strategy to ensure that the rapid convergence
properties associated with Newton’s methods are not lost in the inexact version. In addition,
as noted below, we can implement these methods in a Hessian-free manner, so that the
Hessian ∇ 2 f k need not be calculated or stored explicitly at all.
We examine first how the inexactness in the step calculation determines the local
convergence properties of inexact Newton methods. We then consider line search and
trust-region approaches based on using CG (possibly with preconditioning) to obtain an
approximate solution of (7.1). Finally, we discuss the use of the Lanczos method for solving
(7.1) approximately.
LOCAL CONVERGENCE OF INEXACT NEWTON METHODS
Most rules for terminating the iterative solver for (7.1) are based on the residual
r k ∇ 2 f k pk + ∇ f k ,
(7.2)
where pk is the inexact Newton step. Usually, we terminate the CG iterations when
rk ≤ ηk ∇ f k ,
(7.3)
where the sequence {ηk } (with 0 < ηk < 1 for all k) is called the forcing sequence.
We now study how the rate of convergence of inexact Newton methods based on
(7.1)–(7.3) is affected by the choice of the forcing sequence. The next two theorems apply
not just to Newton–CG procedures but to all inexact Newton methods whose steps satisfy
(7.2) and (7.3).
Our first result says that local convergence is obtained simply by ensuring that ηk is
bounded away from 1.
Theorem 7.1.
Suppose that ∇ 2 f (x) exists and is continuous in a neighborhood of a minimizer x ∗ , with
2
∇ f (x ∗ ) is positive definite. Consider the iteration xk+1 xk + pk where pk satisfies (7.3), and
assume that ηk ≤ η for some constant η ∈ [0, 1). Then, if the starting point x0 is sufficiently
near x ∗ , the sequence {xk } converges to x ∗ and satisfies
∇ 2 f (x ∗ )(xk+1 − x ∗ ) ≤ η̂∇ 2 f (x ∗ )(xk − x ∗ ),
for some constant η̂ with η < η̂ < 1.
(7.4)
7.1.
INEXACT NEWTON METHODS
Rather than giving a rigorous proof of this theorem, we present an informal derivation
that contains the essence of the argument and motivates the next result.
Since the Hessian matrix ∇ 2 f is positive definite at x ∗ and continuous near x ∗ , there
exists a positive constant L such that (∇ 2 f k )−1 ≤ L for all xk sufficiently close to x ∗ . We
therefore have from (7.2) that the inexact Newton step satisfies
pk ≤ L(∇ f k + rk ) ≤ 2L∇ f k ,
where the second inequality follows from (7.3) and ηk < 1. Using this expression together
with Taylor’s theorem and the continuity of ∇ 2 f (x), we obtain
∇ f k+1 ∇ f k + ∇ 2 f k pk +
1
[∇ f (xk + t pk ) − ∇ f (xk )] pk dt
0
∇ f k + ∇ 2 f k pk + o ( pk )
∇ f k − (∇ f k − rk ) + o (∇ f k )
rk + o (∇ f k ) .
(7.5)
Taking norms and recalling (7.3), we have that
∇ f k+1 ≤ ηk ∇ f k + o (∇ f k ) ≤ (ηk + o(1)))∇ f k .
(7.6)
When xk is close enough to x ∗ that the o(1) term in the last estimate is bounded by (1−η)/2,
we have
∇ f k+1 ≤ (ηk + (1 − η)/2)∇ f k ≤
1+η
∇ f k ,
2
(7.7)
so the gradient norm decreases by a factor of (1 + η)/2 at this iteration. By choosing the
initial point x0 sufficiently close to x ∗ , we can ensure that this rate of decrease occurs at
every iteration.
To prove (7.4), we note that under our smoothness assumptions, we have
∇ f k ∇ 2 f (x ∗ )(xk − x ∗ ) + o(xk − x ∗ ).
Hence it can be shown that for xk close to x ∗ , the gradient ∇ f k differs from the scaled error
∇ 2 f (x ∗ )(xk − x ∗ ) by only a relatively small perturbation. A similar estimate holds at xk+1 ,
so (7.4) follows from (7.7).
From (7.6), we have that
∇ f k+1 ≤ ηk + o(1).
∇ f k (7.8)
167
168
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
If limk→∞ ηk 0, we have from this expression that
lim
k→∞
∇ f k+1 0,
∇ f k indicating Q-superlinear convergence of the gradient norms ∇ f k to zero. Superlinear
convergence of the iterates {xk } to x ∗ can be proved as a consequence.
We can obtain quadratic convergence by making the additional assumption that the
Hessian ∇ 2 f (x) is Lipschitz continuous near x ∗ . In this case, the estimate (7.5) can be
tightened to
∇ f k+1 rk + O ∇ f k 2 .
By choosing the forcing sequence so that ηk O(∇ f k ), we have from this expression
that
∇ f k+1 O(∇ f k 2 ),
indicating Q-quadratic convergence of the gradient norms to zero (and thus also Q-quadratic
convergence of the iterates xk to x ∗ ). The last two observations are summarized in the
following theorem.
Theorem 7.2.
Suppose that the conditions of Theorem 7.1 hold, and assume that the iterates {xk }
generated by the inexact Newton method converge to x ∗ . Then the rate of convergence is
superlinear if ηk → 0. If in addition, ∇ 2 f (x) is Lipschitz continuous for x near x ∗ and if
ηk O(∇ f k ), then the convergence is quadratic.
√
To obtain superlinear convergence, we can set, for example, ηk min 0.5, ∇ f k ;
the choice ηk min(0.5, ∇ f k ) would yield quadratic convergence.
All the results presented in this section, which are proved by Dembo, Eisenstat, and
Steihaug [89], are local in nature: They assume that the sequence {xk } eventually enters
the near vicinity of the solution x ∗ . They also assume that the unit step length αk 1 is
taken and hence that globalization strategies do not interfere with rapid convergence. In
the following pages we show that inexact Newton strategies can, in fact, be incorporated
in practical line search and trust-region implementations of Newton’s method, yielding
algorithms with good local and global convergence properties. We start with a line search
approach.
LINE SEARCH NEWTON–CG METHOD
In the line search Newton–CG method, also known as the truncated Newton method, we
compute the search direction by applying the CG method to the Newton equations (7.1) and
7.1.
INEXACT NEWTON METHODS
attempt to satisfy a termination test of the form (7.3). However, the CG method is designed
to solve positive definite systems, and the Hessian ∇ 2 f k may have negative eigenvalues when
xk is not close to a solution. Therefore, we terminate the CG iteration as soon as a direction
of negative curvature is generated. This adaptation of the CG method produces a search
direction pk that is a descent direction. Moreover, the adaptation guarantees that the fast
convergence rate of the pure Newton method is preserved, provided that the step length
αk 1 is used whenever it satisfies the acceptance criteria.
We now describe Algorithm 7.1, a line search algorithm that uses a modification of
Algorithm 5.2 as the inner iteration to compute each search direction pk . For purposes of
this algorithm, we write the linear system (7.1) in the form
Bk p −∇ f k ,
(7.9)
where Bk represents ∇ 2 f k . For the inner CG iteration, we denote the search directions by
d j and the sequence of iterates that it generates by z j . When Bk is positive definite, the
inner iteration sequence {z j } will converge to the Newton step pkN that solves (7.9). At each
major iteration, we define a tolerance k that specifies the required accuracy of the computed
√
solution. For concreteness, we choose the forcing sequence to be ηk min(0.5, ∇ f k )
to obtain a superlinear convergence rate, but other choices are possible.
Algorithm 7.1 (Line Search Newton–CG).
Given initial point x0 ;
for k 0, 1, 2, . . .
√
Define tolerance k min(0.5, ∇ f k )∇ f k ;
Set z 0 0, r0 ∇ f k , d0 −r0 −∇ f k ;
for j 0, 1, 2, . . .
if d Tj Bk d j ≤ 0
if j 0
return pk −∇ f k ;
else
return pk z j ;
Set α j r Tj r j /d Tj Bk d j ;
Set z j+1 z j + α j d j ;
Set r j+1 r j + α j Bk d j ;
if r j+1 < k
return pk z j+1 ;
Set β j+1 r Tj+1r j+1 /r Tj r j ;
Set d j+1 −r j+1 + β j+1 d j ;
end (for)
Set xk+1 xk + αk pk , where αk satisfies the Wolfe, Goldstein, or
Armijo backtracking conditions (using αk 1 if possible);
end
169
170
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
The main differences between the inner loop of Algorithm 7.1 and Algorithm 5.2 are
that the specific starting point z 0 0 is used; the use of a positive tolerance k allows the CG
iterations to terminate at an inexact solution; and the negative curvature test d Tj Bk d j ≤ 0
ensures that pk is a descent direction for f at xk . If negative curvature is detected on the
first inner iteration j 0, the returned direction pk −∇ f k is both a descent direction
and a direction of nonpositive curvature for f at xk .
We can modify the CG iterations in Algorithm 7.1 by introducing preconditioning,
in the manner described in Chapter 5.
Algorithm 7.1 is well suited for large problems, but it has a weakness. When the Hessian
∇ 2 f k is nearly singular, the line search Newton–CG direction can be long and of poor quality,
requiring many function evaluations in the line search and giving only a small reduction in
the function. To alleviate this difficulty, we can try to normalize the Newton step, but good
rules for doing so are difficult to determine. They run the risk of undermining the rapid
convergence of Newton’s method in the case where the pure Newton step is well scaled. It
is preferable to introduce a threshold value into the test d Tj Bd j ≤ 0, but good choices of
the threshold are difficult to determine. The trust-region Newton–CG method described
below deals more effectively with this problematic situation and is therefore preferable, in
our opinion.
The line search Newton–CG method does not require explicit knowledge of the
Hessian Bk ∇ 2 f k . Rather, it requires only that we can supply Hessian–vector products
of the form ∇ 2 f k d for any given vector d. When the user cannot easily supply code to
calculate second derivatives, or where the Hessian requires too much storage, the techniques
of Chapter 8 (automatic differentiation and finite differencing) can be used to calculate these
Hessian–vector products. Methods of this type are known as Hessian-free Newton methods.
To illustrate the finite-differencing technique briefly, we use the approximation
∇ 2 fk d ≈
∇ f (xk + hd) − ∇ f (xk )
,
h
(7.10)
for some small differencing interval h. It is easy to prove that the accuracy of this approximation is O(h); appropriate choices of h are discussed in Chapter 8. The price we pay for
bypassing the computation of the Hessian is one new gradient evaluation per CG iteration.
TRUST-REGION NEWTON–CG METHOD
In Chapter 4, we discussed approaches for finding an approximate solution of the
trust-region subproblem (4.3) that produce improvements on the Cauchy point. Here we
define a modified CG algorithm for solving the subproblem with these properties. This
algorithm, due to Steihaug [281], is specified below as Algorithm 7.2. A complete algorithm
for minimizing f is obtained by using Algorithm 7.2 to generate the step pk required by
Algorithm 4.1 of Chapter 4, for some choice of tolerance k at each iteration.
7.1.
INEXACT NEWTON METHODS
We use notation similar to (7.9) to define the trust-region subproblem for which
Steihaug’s method finds an approximate solution:
def
minn m k ( p) f k + (∇ f k )T p + 12 p T Bk p
p∈IR
subject to p ≤ k ,
(7.11)
where Bk ∇ 2 f k . As in Algorithm 7.1, we use d j to denote the search directions of this
modified CG iteration and z j to denote the sequence of iterates that it generates.
Algorithm 7.2 (CG–Steihaug).
Given tolerance k > 0;
Set z 0 0, r0 ∇ f k , d0 −r0 −∇ f k ;
if r0 < k
return pk z 0 0;
for j 0, 1, 2, . . .
if d Tj Bk d j ≤ 0
Find τ such that pk z j + τ d j minimizes m k ( pk ) in (4.5)
and satisfies pk k ;
return pk ;
Set α j r Tj r j /d Tj Bk d j ;
Set z j+1 z j + α j d j ;
if z j+1 ≥ k
Find τ ≥ 0 such that pk z j + τ d j satisfies pk k ;
return pk ;
Set r j+1 r j + α j Bk d j ;
if r j+1 < k
return pk z j+1 ;
Set β j+1 r Tj+1r j+1 /r Tj r j ;
Set d j+1 −r j+1 + β j+1 d j ;
end (for).
The first if statement inside the loop stops the method if its current search direction
d j is a direction of nonpositive curvature along Bk , while the second if statement inside the
loop causes termination if z j+1 violates the trust-region bound. In both cases, the method
returns the step pk obtained by intersecting the current search direction with the trust-region
boundary.
The choice of the tolerance k at each call to Algorithm 7.2 is important in keeping the
overall cost of the trust-region Newton–CG method low. Near a well-behaved solution x ∗ ,
the trust-region bound becomes inactive, and the method reduces to the inexact Newton
method analyzed in Theorems 7.1 and 7.2. Rapid convergence can be obtained in these
circumstances by choosing k in a similar fashion to Algorithm 7.1.
171
172
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
The essential differences between Algorithm 5.2 and the inner loop of Algorithm 7.2
are that the latter terminates when it violates the trust-region bound p ≤ , when it
encounters a direction of negative curvature in ∇ 2 f k , or when it satisfies a convergence
tolerance defined by a parameter k . In these respects, Algorithm 7.2 is quite similar to the
inner loop of Algorithm 7.1.
The initialization of z 0 to zero in Algorithm 7.2 is a crucial feature of the algorithm.
Provided ∇ f k 2 ≥ k , Algorithm 7.2 terminates at a point pk for which m k ( pk ) ≤ m k ( pkC ),
that is, when the reduction in model function equals or exceeds that of the Cauchy point.
To demonstrate this fact, we consider several cases. First, if d0T Bk d0 (∇ f k )T Bk ∇ f k ≤ 0,
then the condition in the first if statement is satisfied, and the algorithm returns the Cauchy
point p −k (∇ f k )/∇ f k . Otherwise, Algorithm 7.2 defines z 1 as follows:
z 1 α0 d0 r0T r0
(∇ f k )T ∇ f k
d
−
∇ fk .
0
(∇ f k )T Bk ∇ f k
d0T Bk d0
If z 1 < k , then z 1 is exactly the Cauchy point. Subsequent steps of Algorithm 7.2 ensure
that the final pk satisfies m k ( pk ) ≤ m k (z 1 ). When z 1 ≥ k , on the other hand, the second
if statement is activated, and Algorithm 7.2 terminates at the Cauchy point, proving our
claim. This property is important for global convergence: Since each step is at least as good
as the Cauchy point in reducing the model m k , Algorithm 7.2 is globally convergent.
Another crucial property of the method is that each iterate z j is larger in norm than
its predecessor. This property is another consequence of the initialization z 0 0. Its main
implication is that it is acceptable to stop iterating as soon as the trust-region boundary is
reached, because no further iterates giving a lower value of the model function m k will lie
inside the trust region. We state and prove this property formally in the following theorem,
which makes use of the expanding subspace property of the conjugate gradient algorithm,
described in Theorem 5.2.
Theorem 7.3.
The sequence of vectors {z j } generated by Algorithm 7.2 satisfies
0 z 0 2 < · · · < z j 2 < z j+1 2 < · · · < pk 2 ≤ k .
PROOF. We first show that the sequences of vectors generated by Algorithm 7.2 satisfy
z Tj r j 0 for j ≥ 0 and z Tj d j > 0 for j ≥ 1.
Algorithm 7.2 computes z j+1 recursively in terms of z j ; but when all the terms of this
recursion are written explicitly, we see that
z j z0 +
j−1
i0
αi di j−1
i0
αi di ,
7.1.
INEXACT NEWTON METHODS
since z 0 0. Multiplying by r j and applying the expanding subspace property of conjugate
gradients (see Theorem 5.2), we obtain
z Tj r j
j−1
αi diT r j 0.
(7.12)
i0
An induction proof establishes the relation z Tj d j > 0. By applying the expanding
subspace property again, we obtain
z 1T d1 (α0 d0 )T (−r1 + β1 d0 ) α0 β1 d0T d0 > 0.
We now make the inductive hypothesis that z Tj d j > 0 and deduce that z Tj+1 d j+1 > 0. From
(7.12), we have z Tj+1r j+1 0, and therefore
z Tj+1 d j+1 z Tj+1 (−r j+1 + β j+1 d j )
β j+1 z Tj+1 d j
β j+1 (z j + α j d j )T d j
β j+1 z Tj d j + α j β j+1 d Tj d j .
Because of the inductive hypothesis and positivity of β j+1 and α j , the last expression is
positive.
We now prove the theorem. If Algorithm 7.2 terminates because d Tj Bk d j ≤ 0 or
z j+1 2 ≥ k , then the final point pk is chosen to make pk 2 k , which is the
largest possible length. To cover all other possibilities in the algorithm, we must show that
z j 2 < z j+1 2 when z j+1 z j + α j d j and j ≥ 1. Observe that
z j+1 22 (z j + α j d j )T (z j + α j d j ) z j 22 + 2α j z Tj d j + α 2j d j 22 .
It follows from this expression and our intermediate result that z j 2 < z j+1 2 , so our
proof is complete.
From this theorem we see that Algorithm 7.2 sweeps out points z j that move on some
interpolating path from z 1 to the final solution pk , a path in which every step increases its
total distance from the start point. When Bk ∇ 2 f k is positive definite, this path may
be compared to the path of the dogleg method: Both methods start by minimizing m k
along the negative gradient direction −∇ f k and subsequently progress toward pkN , until the
trust-region boundary intervenes. One can show that, when Bk ∇ 2 f k is positive definite,
Algorithm 7.2 provides a decrease in the model (7.11) that is at least half as good as the
optimal decrease [320].
173
174
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
PRECONDITIONING THE TRUST-REGION NEWTON–CG METHOD
As discussed in Chapter 5, preconditioning can be used to accelerate the CG iteration.
Preconditioning techniques are based on finding a nonsingular matrix D such that the eigenvalues of D −T ∇ 2 f k D −1 have a more favorable distribution. By generalizing Theorem 7.3,
we can show that the iterates z j generated by a preconditioned variant of Algorithm 7.2 will
grow monotonically in the weighted norm D · . To be consistent, we should redefine the
trust-region subproblem in terms of the same norm, as follows:
def
minn m k ( p) f k + ∇ f k T p + 12 p T Bk p
p∈IR
subject to Dp ≤ k .
(7.13)
Making the change of variables p̂ Dp and defining
ĝk D −T ∇ f k ,
B̂k D −T (∇ 2 f k )D −1 ,
we can write (7.13) as
minn f k + ĝkT p̂ + 12 p̂ T B̂k p̂
p̂∈IR
subject to p̂ ≤ ,
which has exactly the form of (7.11). We can apply Algorithm 7.2 without any modification to
this subproblem, which is equivalent to applying a preconditioned version of Algorithm 7.2
to the problem (7.13).
Many preconditioners can be used within this framework; we discuss some of them
in Chapter 5. Of particular interest is incomplete Cholesky factorization, which has proved
useful in a wide range of optimization problems. The incomplete Cholesky factorization of
a positive definite matrix B finds a lower triangular matrix L such that
B L L T − R,
where the amount of fill-in in L is restricted in some way. (For instance, it is constrained
to have the same sparsity structure as the lower triangular part of B or is allowed to have a
number of nonzero entries similar to that in B.) The matrix R accounts for the inexactness
in the approximate factorization. The situation is complicated somewhat by the possible
indefiniteness of the Hessian ∇ 2 f k ; we must be able to handle this indefiniteness as well as
maintain the sparsity. The following algorithm combines incomplete Cholesky and a form
of modified Cholesky to define a preconditioner for the trust-region Newton–CG approach.
Algorithm 7.3 (Inexact Modified Cholesky).
Compute T diag(Be1 , Be2 , . . . , Ben ), where ei is the
ith coordinate vector;
Set B̄ ← T −1/2 BT −1/2 ; Set β ← B̄;
7.1.
INEXACT NEWTON METHODS
(compute a shift to ensure positive definiteness)
if mini bii > 0
α0 ← 0
else
α0 ← β/2;
for k 0, 1, 2, . . .
Attempt to apply incomplete Cholesky algorithm to obtain
L L T B̄ + αk I ;
if the factorization is completed successfully
stop and return L;
else
αk+1 ← max(2αk , β/2);
end (for)
We can then set the preconditioner to be D L T , where L is the lower triangular matrix
output from Algorithm 7.3. A trust-region Newton–CG method using this preconditioner
is implemented in the LANCELOT [72] and TRON [192] codes.
TRUST-REGION NEWTON–LANCZOS METHOD
A limitation of Algorithm 7.2 is that it accepts any direction of negative curvature, even
when this direction gives an insignificant reduction in the model. Consider, for example,
the case where the subproblem (7.11) is
min m( p) 10−3 p1 − 10−4 p12 − p22
p
subject to p ≤ 1,
where subscripts indicate elements of the vector p. The steepest descent direction at p 0 is
(−10−3 , 0)T , which is a direction of negative curvature for the model. Algorithm 7.2 would
follow this direction to the boundary of the trust region, yielding a reduction in model
function m of about 10−3 . A step along e2 —also a direction of negative curvature—would
yield a much greater reduction of 1.
Several remedies have been proposed. We have seen in Chapter 4 that when the
Hessian ∇ 2 f k contains negative eigenvalues, the search direction should have a significant
component along the eigenvector corresponding to the most negative eigenvalue of ∇ 2 f k .
This feature would allow the algorithm to move away rapidly from stationary points that
are not minimizers. One way to achieve this is to compute a nearly exact solution of
the trust-region subproblem (7.11) using the techniques described in Section 4.3. This
approach requires the solution of a few linear systems with coefficient matrices of the form
175
176
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
Bk + λI . Although this approach is perhaps too expensive in the large-scale case, it generates
productive search directions in all cases.
A more practical alternative is to use the Lanczos method (see, for example, [136])
rather than the CG method to solve the linear system Bk p −∇ f k . The Lanczos method
can be seen as a generalization of the CG method that is applicable to indefinite systems, and
we can use it to continue the CG process while gathering negative curvature information.
After j steps, the Lanczos method generates an n × j matrix Q j with orthogonal
columns that span the Krylov subspace (5.15) generated by this method. This matrix has
the property that Q Tj B Q j T j , where T j is an tridiagonal. We can take advantage of
this tridiagonal structure and seek to find an approximate solution of the trust-region
subproblem in the range of the basis Q j . To do so, we solve the problem
min f k + e1T Q j (∇ f k )e1T w + 12 w T T j w
w∈IR j
subject to w ≤ k ,
(7.14)
where e1 (1, 0, 0, . . . , 0)T , and we define the approximate solution of the trust-region
subproblem as pk Q j w. Since T j is tridiagonal, problem (7.14) can be solved by factoring
the system T j + λI and following the (nearly) exact approach of Section 4.3.
The Lanczos iteration may be terminated, as in the Newton–CG methods, by a test of
the form (7.3). Preconditioning can also be incorporated to accelerate the convergence of
the Lanczos iteration. The additional robustness in this trust-region algorithm comes at the
cost of a more expensive solution of the subproblem than in the Newton–CG approach. A
sophisticated implementation of the Newton–Lanczos approach has been implemented in
the GLTR package [145].
7.2
LIMITED-MEMORY QUASI-NEWTON METHODS
Limited-memory quasi-Newton methods are useful for solving large problems whose Hessian matrices cannot be computed at a reasonable cost or are not sparse. These methods
maintain simple and compact approximations of Hessian matrices: Instead of storing fully
dense n × n approximations, they save only a few vectors of length n that represent the
approximations implicitly. Despite these modest storage requirements, they often yield an
acceptable (albeit linear) rate of convergence. Various limited-memory methods have been
proposed; we focus mainly on an algorithm known as L-BFGS, which, as its name suggests,
is based on the BFGS updating formula. The main idea of this method is to use curvature
information from only the most recent iterations to construct the Hessian approximation.
Curvature information from earlier iterations, which is less likely to be relevant to the actual behavior of the Hessian at the current iteration, is discarded in the interest of saving
storage.
Following our discussion of L-BFGS and its convergence behavior, we discuss its
relationship to the nonlinear conjugate gradient methods of Chapter 5. We then discuss
7.2.
LIMITED-MEMORY QUASI-NEWTON METHODS
implementations of limited-memory schemes that make use of a compact representation of
approximate Hessian information. These techniques can be applied not only to L-BFGS but
also to limited-memory versions of other quasi-Newton procedures such as SR1. Finally,
we discuss quasi-Newton updating schemes that impose a particular sparsity pattern on the
approximate Hessian.
LIMITED-MEMORY BFGS
We begin our description of the L-BFGS method by recalling its parent, the BFGS
method, which was described in Algorithm 8.1. Each step of the BFGS method has the form
xk+1 xk − αk Hk ∇ f k ,
(7.15)
where αk is the step length and Hk is updated at every iteration by means of the formula
Hk+1 VkT Hk Vk + ρk sk skT
(7.16)
(see (6.17)), where
ρk 1
,
ykT sk
Vk I − ρk yk skT ,
(7.17)
and
sk xk+1 − xk ,
yk ∇ f k+1 − ∇ f k .
(7.18)
Since the inverse Hessian approximation Hk will generally be dense, the cost of storing
and manipulating it is prohibitive when the number of variables is large. To circumvent this
problem, we store a modified version of Hk implicitly, by storing a certain number (say, m)
of the vector pairs {si , yi } used in the formulas (7.16)–(7.18). The product Hk ∇ f k can be
obtained by performing a sequence of inner products and vector summations involving
∇ f k and the pairs {si , yi }. After the new iterate is computed, the oldest vector pair in the set
of pairs {si , yi } is replaced by the new pair {sk , yk } obtained from the current step (7.18).
In this way, the set of vector pairs includes curvature information from the m most recent
iterations. Practical experience has shown that modest values of m (between 3 and 20, say)
often produce satisfactory results.
We now describe the updating process in a little more detail. At iteration k, the current
iterate is xk and the set of vector pairs is given by {si , yi } for i k − m, . . . , k − 1. We first
choose some initial Hessian approximation Hk0 (in contrast to the standard BFGS iteration,
this initial approximation is allowed to vary from iteration to iteration) and find by repeated
application of the formula (7.16) that the L-BFGS approximation Hk satisfies the following
177
178
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
formula:
0
T
T
Hk Vk−1
Hk (Vk−m · · · Vk−1 )
· · · Vk−m
T
T
T
sk−m sk−m
+ ρk−m Vk−1 · · · Vk−m+1
(Vk−m+1 · · · Vk−1 )
T
T
T
+ ρk−m+1 Vk−1 · · · Vk−m+2 sk−m+1 sk−m+1
(Vk−m+2 · · · Vk−1 )
+ ···
T
.
+ ρk−1 sk−1 sk−1
(7.19)
From this expression we can derive a recursive procedure to compute the product Hk ∇ f k
efficiently.
Algorithm 7.4 (L-BFGS two-loop recursion).
q ← ∇ fk ;
for i k − 1, k − 2, . . . , k − m
αi ← ρi siT q;
q ← q − αi yi ;
end (for)
r ← Hk0 q;
for i k − m, k − m + 1, . . . , k − 1
β ← ρi yiT r ;
r ← r + si (αi − β)
end (for)
stop with result Hk ∇ f k r .
Without considering the multiplication Hk0 q, the two-loop recursion scheme requires
4mn multiplications; if Hk0 is diagonal, then n additional multiplications are needed. Apart
from being inexpensive, this recursion has the advantage that the multiplication by the
initial matrix Hk0 is isolated from the rest of the computations, allowing this matrix to be
chosen freely and to vary between iterations. We may even use an implicit choice of Hk0 by
defining some initial approximation Bk0 to the Hessian (not its inverse) and obtaining r by
solving the system Bk0r q.
A method for choosing Hk0 that has proved effective in practice is to set Hk0 γk I ,
where
γk T
yk−1
sk−1
T
yk−1
yk−1
.
(7.20)
As discussed in Chapter 6, γk is the scaling factor that attempts to estimate the size of the
true Hessian matrix along the most recent search direction (see (6.21)). This choice helps
to ensure that the search direction pk is well scaled, and as a result the step length αk 1 is
accepted in most iterations. As discussed in Chapter 6, it is important that the line search be
7.2.
LIMITED-MEMORY QUASI-NEWTON METHODS
based on the Wolfe conditions (3.6) or strong Wolfe conditions (3.7), so that BFGS updating
is stable.
The limited-memory BFGS algorithm can be stated formally as follows.
Algorithm 7.5 (L-BFGS).
Choose starting point x0 , integer m > 0;
k ← 0;
repeat
Choose Hk0 (for example, by using (7.20));
Compute pk ← −Hk ∇ f k from Algorithm 7.4;
Compute xk+1 ← xk + αk pk , where αk is chosen to
satisfy the Wolfe conditions;
if k > m
Discard the vector pair {sk−m , yk−m } from storage;
Compute and save sk ← xk+1 − xk , yk ∇ f k+1 − ∇ f k ;
k ← k + 1;
until convergence.
The strategy of keeping the m most recent correction pairs {si , yi } works well in
practice; indeed no other strategy has yet proved to be consistently better. During its first
m − 1 iterations, Algorithm 7.5 is equivalent to the BFGS algorithm of Chapter 6 if the
initial matrix H0 is the same in both methods, and if L-BFGS chooses Hk0 H0 at each
iteration.
Table 7.1 presents results illustrating the behavior of Algorithm 7.5 for various levels
of memory m. It gives the number of function and gradient evaluations (nfg) and the total
CPU time. The test problems are taken from the CUTE collection [35], the number of
variables is indicated by n, and the termination criterion ∇ f k ≤ 10−5 is used. The table
shows that the algorithm tends to be less robust when m is small. As the amount of storage
increases, the number of function evaluations tends to decrease; but since the cost of each
iteration increases with the amount of storage, the best CPU time is often obtained for small
values of m. Clearly, the optimal choice of m is problem dependent.
Because some rival algorithms are inefficient, Algorithm 7.5 is often the approach of
choice for large problems in which the true Hessian is not sparse. In particular, a Newton
Table 7.1 Performance of Algorithm 7.5.
Problem
DIXMAANL
EIGENALS
FREUROTH
TRIDIA
n
1500
110
1000
1000
L-BFGS
m3
nfg time
146
821
>999
876
16.5
21.5
—
46.6
L-BFGS
m5
nfg time
134
569
>999
611
17.4
15.7
—
41.4
L-BFGS
m 17
nfg time
L-BFGS
m 29
nfg
time
120
363
69
531
125
168
38
462
28.2
16.2
8.1
84.6
44.4
12.5
6.3
127.1
179
180
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
method in which the exact Hessian is computed and factorized is not practical in such
circumstances. The L-BFGS approach may also outperform Hessian-free Newton methods
such as Newton–CG approaches, in which Hessian–vector products are calculated by finite
differences or automatic differentiation. The main weakness of the L-BFGS method is that it
converges slowly on ill-conditioned problems—specifically, on problems where the Hessian
matrix contains a wide distribution of eigenvalues. On certain applications, the nonlinear
conjugate gradient methods discussed in Chapter 5 are competitive with limited-memory
quasi-Newton methods.
RELATIONSHIP WITH CONJUGATE GRADIENT METHODS
Limited-memory methods evolved as an attempt to improve nonlinear conjugate
gradient methods, and early implementations resembled conjugate gradient methods more
than quasi-Newton methods. The relationship between the two classes is the basis of a
memoryless BFGS iteration, which we now outline.
We start by considering the Hestenes–Stiefel form of the nonlinear conjugate gradient
method (5.46). Recalling that sk αk pk , we have that the search direction for this method
is given by
pk+1 −∇ f k+1 +
T
yk
∇ f k+1
ykT pk
sk y T
pk − I − T k
yk sk
∇ f k+1 ≡ − Ĥk+1 ∇ f k+1 .
(7.21)
This formula resembles a quasi-Newton iteration, but the matrix Ĥk+1 is neither symmetric
T
Ĥk+1 , but this matrix does not satisfy
nor positive definite. We could symmetrize it as Ĥk+1
the secant equation Ĥk+1 yk sk and is, in any case, singular. An iteration matrix that is
symmetric, positive definite, and satisfies the secant equation is given by
Hk+1 I−
sk ykT
ykT sk
I−
yk skT
ykT sk
+
sk skT
.
ykT sk
(7.22)
This matrix is exactly the one obtained by applying a single BFGS update (7.16) to the identity
matrix. Hence, an algorithm whose search direction is given by pk+1 −Hk+1 ∇ f k+1 , with
Hk+1 defined by (7.22), can be thought of as a “memoryless” BFGS method, in which the
previous Hessian approximation is always reset to the identity matrix before updating it and
where only the most recent correction pair (sk , yk ) is kept at every iteration. Alternatively,
we can view the method as a variant of Algorithm 7.5 in which m 1 and Hk0 I at each
iteration.
A more direct connection with conjugate gradient methods can be seen if we consider
the memoryless BFGS formula (7.22) in conjunction with an exact line search, for which
7.2.
LIMITED-MEMORY QUASI-NEWTON METHODS
T
∇ f k+1
pk 0 for all k. We then obtain
pk+1 −Hk+1 ∇ f k+1 −∇ f k+1 +
T
yk
∇ f k+1
ykT pk
pk ,
(7.23)
which is none other than the Hestenes–Stiefel conjugate gradient method. Moreover, it is
T
pk 0, the Hestenes–Stiefel formula reduces to the Polak–
easy to verify that when ∇ f k+1
Ribière formula (5.44). Even though the assumption of exact line searches is unrealistic,
it is intriguing that the BFGS formula is related in this way to the Polak–Ribière and
Hestenes–Stiefel methods.
GENERAL LIMITED-MEMORY UPDATING
Limited-memory quasi-Newton approximations are useful in a variety of optimization
methods. L-BFGS, Algorithm 7.5, is a line search method for unconstrained optimization
that (implicitly) updates an approximation Hk to the inverse of the Hessian matrix. Trustregion methods, on the other hand, require an approximation Bk to the Hessian matrix,
not to its inverse. We would also like to develop limited-memory methods based on the SR1
formula, which is an attractive alternative to BFGS; see Chapter 6. In this section we consider
limited-memory updating in a general setting and show that by representing quasi-Newton
matrices in a compact (or outer product) form, we can derive efficient implementations of all
popular quasi-Newton update formulas, and their inverses. These compact representations
will also be useful in designing limited-memory methods for constrained optimization,
where approximations to the Hessian or reduced Hessian of the Lagrangian are needed; see
Chapter 18 and Chapter 19.
We will consider only limited-memory methods (such as L-BFGS) that continuously
refresh the correction pairs by removing and adding information at each stage. A different
approach saves correction pairs until the available storage is exhausted and then discards all
correction pairs (except perhaps one) and starts the process anew. Computational experience
suggests that this second approach is less effective in practice.
Throughout this chapter we let Bk denote an approximation to a Hessian matrix and
Hk the approximation to the inverse. In particular, we always have that Bk−1 Hk .
COMPACT REPRESENTATION OF BFGS UPDATING
We now describe an approach to limited-memory updating that is based on representing quasi-Newton matrices in outer-product form. We illustrate it for the case of a BFGS
approximation Bk to the Hessian.
Theorem 7.4.
k−1
Let B0 be symmetric and positive definite, and assume that the k vector pairs {si , yi }i0
T
satisfy si yi > 0. Let Bk be obtained by applying k BFGS updates with these vector pairs to B0 ,
181
182
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
using the formula (6.19). We then have that
Bk B0 −
B0 Sk
Yk
SkT B0 Sk
Lk
L kT
−Dk
−1 SkT B0
,
YkT
(7.24)
where Sk and Yk are the n × k matrices defined by
Sk [s0 , . . . , sk−1 ] ,
Yk [y0 , . . . , yk−1 ] ,
(7.25)
while L k and Dk are the k × k matrices
(L k )i, j T
si−1
y j−1
if i > j,
0
otherwise,
(7.26)
T
Dk diag s0T y0 , . . . , sk−1
yk−1 .
(7.27)
This result can be proved by induction. We note that the conditions siT yi > 0, i 0, 1, . . . , k − 1, ensure that the middle matrix in (7.24) is nonsingular, so that this expression is well defined. The utility of this representation becomes apparent when we consider
limited-memory updating.
As in the L-BFGS algorithm, we keep the m most recent correction pairs {si , yi } and
refresh this set at every iteration by removing the oldest pair and adding a newly generated
pair. During the first m iterations, the update procedure described in Theorem 7.4 can be
used without modification, except that usually we make the specific choice Bk0 δk I for
the basic matrix, where δk 1/γk and γk is defined by (7.20).
At subsequent iterations k > m, the update procedure needs to be modified slightly to
reflect the changing nature of the set of vector pairs {si , yi } for i k−m, k−m+1, . . . , k−1.
Defining the n × m matrices Sk and Yk by
Sk [sk−m , . . . , sk−1 ] ,
Yk [yk−m , . . . , yk−1 ] ,
(7.28)
we find that the matrix Bk resulting from m updates to the basic matrix B0(k) δk I is given
by
Bk δk I −
δk Sk
Yk
δk SkT Sk
Lk
L kT
−Dk
−1 δk SkT
YkT
where L k and Dk are now the m × m matrices defined by
(L k )i, j (sk−m−1+i )T (yk−m−1+ j )
if i > j,
0
otherwise,
Dk diag
T
T
sk−m
yk−m , . . . , sk−1
yk−1
.
,
(7.29)
7.2.
Bk = δk I+
LIMITED-MEMORY QUASI-NEWTON METHODS
Figure 7.1
Compact (or outer
product) representation of
Bk in (7.29).
After the new iterate xk+1 is generated, we obtain Sk+1 by deleting sk−m from Sk and adding
the new displacement sk , and we update Yk+1 in a similar fashion. The new matrices L k+1
and Dk+1 are obtained in an analogous way.
Since the middle matrix in (7.29) is small—of dimension 2m—its factorization requires a negligible amount of computation. The key idea behind the compact representation
(7.29) is that the corrections to the basic matrix can be expressed as an outer product of two
long and narrow matrices—[δk Sk Yk ] and its transpose—with an intervening multiplication
by a small 2m × 2m matrix. See Figure 7.1 for a graphical illustration.
The limited-memory updating procedure of Bk requires approximately 2mn + O(m 3 )
operations, and matrix–vector products of the form Bk v can be performed at a cost of
(4m + 1)n + O(m 2 ) multiplications. These operation counts indicate that updating and
manipulating the direct limited-memory BFGS matrix Bk is quite economical when m is
small.
This approximation Bk can be used in a trust-region method for unconstrained optimization or, more significantly, in methods for bound-constrained and general-constrained
optimization. The program L-BFGS-B [322] makes extensive use of compact limited-memory
approximations to solve large nonlinear optimization problems with bound constraints. In
this situation, projections of Bk into subspaces defined by the constraint gradients must be
calculated repeatedly. Several codes for general-constrained optimization, including KNITRO
and IPOPT, make use of the compact limited-memory matrix Bk to approximate the Hessian
of the Lagrangians; see Section 19.3
We can derive a formula, similar to (7.24), that provides a compact representation
of the inverse BFGS approximation Hk ; see [52] for details. An implementation of the
unconstrained L-BFGS algorithm based on this expression requires a similar amount of
computation as the algorithm described in the previous section.
Compact representations can also be derived for matrices generated by the symmetric
rank-one (SR1) formula. If k updates are applied to the symmetric matrix B0 using the
k−1
and the SR1 formula (6.24), the resulting matrix Bk can be expressed
vector pairs {si , yi }i0
as
Bk B0 + (Yk − B0 Sk )(Dk + L k + L kT − SkT B0 Sk )−1 (Yk − B0 Sk )T ,
(7.30)
183
184
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
where Sk , Yk , Dk , and L k are as defined in (7.25), (7.26), and (7.27). Since the SR1 method
is self-dual, the inverse formula Hk can be obtained simply by replacing B, s, and y by H ,
y, and s, respectively. Limited-memory SR1 methods can be derived in the same way as the
BFGS method. We replace B0 with the basic matrix Bk0 at the kth iteration, and we redefine
Sk and Yk to contain the m most recent corrections, as in (7.28). We note, however, that
limited-memory SR1 updating is sometimes not as effective as L-BFGS updating because it
may not produce positive definite approximations near a solution.
UNROLLING THE UPDATE
The reader may wonder whether limited-memory updating can be implemented
in simpler ways. In fact, as we show here, the most obvious implementation of limitedmemory BFGS updating is considerably more expensive than the approach based on compact
representations discussed in the previous section.
The direct BFGS formula (6.19) can be written as
Bk+1 Bk − ak akT + bk bkT ,
(7.31)
where the vectors ak and bk are defined by
ak Bk sk
(skT
Bk sk )
1
2
,
bk yk
1
(ykT sk ) 2
.
(7.32)
We could continue to save the vector pairs {si , yi } but use the formula (7.31) to compute
matrix–vector products. A limited-memory BFGS method that uses this approach would
proceed by defining the basic matrix Bk0 at each iteration and then updating according to
the formula
Bk Bk0
+
k−1
bi biT − ai aiT .
(7.33)
ik−m
The vector pairs {ai , bi }, i k − m, k − m + 1, . . . , k − 1, would then be recovered from
the stored vector pairs {si , yi }, i k − m, k − m + 1, . . . , k − 1, by the following procedure:
Procedure 7.6 (Unrolling the BFGS formula).
for i k − m, k − m + 1, . . . , k − 1
bi ← yi /(yiT si )1/2 ;
)
*
T
T
(b
;
s
)b
−
(a
s
)a
ai ← Bk0 si + i−1
i
j
i
j
jk−m
j
j
ai ← ai /(siT ai )1/2 ;
end (for)
7.3.
SPARSE QUASI-NEWTON UPDATES
Note that the vectors ai must be recomputed at each iteration because they all depend
on the vector pair {sk−m , yk−m }, which is removed at the end of iteration k. On the other
hand, the vectors bi and the inner products b Tj si can be saved from the previous iteration,
so only the new values bk−1 and b Tj sk−1 need to be computed at the current iteration.
By taking all these computations into account, and assuming that Bk0 I , we find
that approximately 32 m 2 n operations are needed to determine the limited-memory matrix.
The actual computation of the inner product Bm v (for arbitrary v ∈ IRn ) requires 4mn
multiplications. Overall, therefore, this approach is less efficient than the one based on the
compact matrix representation described previously. Indeed, while the product Bk v costs
the same in both cases, updating the representation of the limited-memory matrix by using
the compact form requires only 2mn multiplications, compared to 32 m 2 n multiplications
needed when the BFGS formula is unrolled.
7.3
SPARSE QUASI-NEWTON UPDATES
We now discuss a quasi-Newton approach to large-scale problems that has intuitive appeal:
We demand that the quasi-Newton approximations Bk have the same (or similar) sparsity
pattern as the true Hessian. This approach would reduce the storage requirements of the
algorithm and perhaps give rise to more accurate Hessian approximations.
Suppose that we know which components of the Hessian may be nonzero at some
point in the domain of interest. That is, we know the contents of the set defined by
def
{(i, j) | [∇ 2 f (x)]i j 0 for some x in the domain of f }.
Suppose also that the current Hessian approximation Bk mirrors the nonzero structure of
/ . In updating Bk to Bk+1 , then, we
the exact Hessian, that is, (Bk )i j 0 for (i, j) ∈
could try to find the matrix Bk+1 that satisfies the secant condition, has the same sparsity
pattern, and is as close as possible to Bk . Specifically, we define Bk+1 to be the solution of
the following quadratic program:
min B − Bk 2F B
[Bi j − (Bk )i j ]2 ,
(7.34a)
(i, j)∈
/ .
subject to Bsk yk , B B T , and Bi j 0 for (i, j) ∈
(7.34b)
One can show that the solution Bk+1 of this problem can be obtained by solving an n × n
linear system whose sparsity pattern is , the same as the sparsity of the true Hessian. Once
Bk+1 has been computed, we can use it, within a trust-region method, to obtain the new
iterate xk+1 . We note that Bk+1 is not guaranteed to be positive definite.
We omit further details of this approach because it has several drawbacks. The updating
process does not possess scale invariance under linear transformations of the variables and,
185
186
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
more significantly, its practical performance has been disappointing. The fundamental
weakness of this approach is that (7.34a) is an inadequate model and can produce poor
Hessian approximations.
An alternative approach is to relax the secant equation, making sure that it is approximately satisfied along the last few steps rather than requiring it to hold strictly on the latest
step. To do so, we define Sk and Yk by (7.28) so that they contain the m most recent difference
pairs. We can then define the new Hessian approximation Bk+1 to be the solution of
min B Sk − Yk 2F
B
subject to B B T and Bi j 0 for (i, j) ∈
/ .
This convex optimization problem has a solution, but it is not easy to compute. Moreover,
this approach can produce singular or poorly conditioned Hessian approximations. Even
though it frequently outperforms methods based on (7.34a), its performance on large
problems has not been impressive.
7.4
ALGORITHMS FOR PARTIALLY SEPARABLE FUNCTIONS
In a separable unconstrained optimization problem, the objective function can be decomposed into a sum of simpler functions that can be optimized independently. For example, if
we have
f (x) f 1 (x1 , x3 ) + f 2 (x2 , x4 , x6 ) + f 3 (x5 ),
we can find the optimal value of x by minimizing each function f i , i 1, 2, 3, independently, since no variable appears in more than one function. The cost of performing m
lower-dimensional optimizations is much less in general than the cost of optimizing an
n-dimensional function.
In many large problems the objective function f : IRn → IR is not separable, but
it can still be written as the sum of simpler functions, known as element functions. Each
element function has the property that it is unaffected when we move along a large number
of linearly independent directions. If this property holds, we say that f is partially separable.
All functions whose Hessians ∇ 2 f are sparse are partially separable, but so are many
functions whose Hessian is not sparse. Partial separability allows for economical problem
representation, efficient automatic differentiation, and effective quasi-Newton updating.
The simplest form of partial separability arises when the objective function can be
written as
f (x) ne
i1
f i (x),
(7.35)
7.4.
ALGORITHMS
FOR
PARTIALLY SEPARABLE FUNCTIONS
where each of the element functions f i depends on only a few components of x. It follows
that the gradients ∇ f i and Hessians ∇ 2 f i of each element function contain just a few
nonzeros. By differentiating (7.35), we obtain
∇ f (x) ne
∇ f i (x),
∇ 2 f (x) i1
ne
∇ 2 f i (x).
i1
A natural question is whether it is more effective to maintain quasi-Newton approximations
to each of the element Hessians ∇ 2 f i (x) separately, rather than approximating the entire
Hessian ∇ 2 f (x). We will show that the answer is affirmative, provided that the quasi-Newton
approximation fully exploits the structure of each element Hessian.
We introduce the concept by means of a simple example. Consider the objective
function
f (x) (x1 − x32 )2 + (x2 − x42 )2 + (x3 − x22 )2 + (x4 − x12 )2
(7.36)
≡ f 1 (x) + f 2 (x) + f 3 (x) + f 4 (x).
The Hessians of the element functions f i are 4 × 4 sparse, singular matrices with 4 nonzero
entries.
Let us focus on f 1 ; all other element functions have exactly the same form. Even
though f 1 is formally a function of all components of x, it depends only on x1 and x3 , which
we call the element variables for f 1 . We assemble the element variables into a vector that we
call x[1] , that is,
x[1] x1
,
x3
and note that
x[1] U1 x
with
U1 1
0
0
0
0
0
1
0
.
If we define the function φ1 by
φ1 (z 1 , z 2 ) (z 1 − z 22 )2 ,
then we can write f 1 (x) φ1 (U1 x). By applying the chain rule to this representation, we
obtain
∇ f 1 (x) U1T ∇φ1 (U1 x),
∇ 2 f 1 (x) U1T ∇ 2 φ1 (U1 x)U1 .
(7.37)
187
188
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
In our case, we have
⎡
∇ φ1 (U1 x) 2
2
−4x3
−4x3
12x32 − 4x1
2
⎢
⎢ 0
∇ f 1 (x) ⎢
⎢ −4x
3
⎣
0
,
2
0
−4x3
0
0
0
12x32 − 4x1
0
0
0
⎤
⎥
0 ⎥
⎥.
0 ⎥
⎦
0
The matrix U1 , known as a compactifying matrix, allows us to map the derivative information
for the low-dimensional function φ1 into the derivative information for the element function
f1 .
Now comes the key idea: Instead of maintaining a quasi-Newton approximation to
∇ 2 f 1 , we maintain a 2 × 2 quasi-Newton approximation B[1] of ∇ 2 φ1 and use the relation
(7.37) to transform it into a quasi-Newton approximation to ∇ 2 f 1 . To update B[1] after a
typical step from x to x + , we record the information
+
− x[1] ,
s[1] x[1]
+
y[1] ∇φ1 (x[1]
) − ∇φ1 (x[1] ),
(7.38)
+
. We therefore update
and use BFGS or SR1 updating to obtain the new approximation B[1]
small, dense quasi-Newton approximations with the property
B[1] ≈ ∇ 2 φ1 (U1 x) ∇ 2 φ1 (x[1] ).
(7.39)
To obtain an approximation of the element Hessian ∇ 2 f 1 , we use the transformation
suggested by the relationship (7.37); that is,
∇ 2 f 1 (x) ≈ U1T B[1] U1 .
This operation has the effect of mapping the elements of B[1] to the correct positions in the
full n × n Hessian approximation.
The previous discussion concerned only the first element function f 1 , but we can treat
all other functions f i in the same way. The full objective function can now be written as
f (x) ne
φi (Ui x),
(7.40)
i1
and we maintain a quasi-Newton approximation B[i] for each of the functions φi . To obtain
a complete approximation to the full Hessian ∇ 2 f , we simply sum the element Hessian
approximations as follows:
B
ne
i1
UiT B[i] Ui .
(7.41)
7.5.
PERSPECTIVES
AND
SOFTWARE
We may use this approximate Hessian in a trust-region algorithm, obtaining an
approximate solution pk of the system
Bk pk −∇ f k .
(7.42)
We need not assemble Bk explicitly but rather use the conjugate gradient approach to solve
(7.42), computing matrix–vector products of the form Bk v by performing operations with
the matrices Ui and B[i] .
To illustrate the usefulness of this element-by-element updating technique, let us
consider a problem of the form (7.36) but this time involving 1000 variables, not just 4. The
functions φi still depend on only two internal variables, so that each Hessian approximation
B[i] is a 2 × 2 matrix. After just a few iterations, we will have sampled enough directions
s[i] to make each B[i] an accurate approximation to ∇ 2 φi . Hence the full quasi-Newton
approximation (7.41) will tend to be a very good approximation to ∇ 2 f (x). By contrast, a
quasi-Newton method that ignores the partially separable structure of the objective function
will attempt to estimate the total average curvature—the sum of the individual curvatures
of the element functions—by approximating the 1000 × 1000 Hessian matrix. When the
number of variables n is large, many iterations will be required before this quasi-Newton
approximation is of good quality. Hence an algorithm of this type (for example, standard
BFGS or L-BFGS) will require many more iterations than a method based on the partially
separable approximate Hessian.
It is not always possible to use the BFGS formula to update the partial Hessian B[i] ,
T
y[i] > 0 will be satisfied. That
because there is no guarantee that the curvature condition s[i]
2
is, even though the full Hessian ∇ f (x) is at least positive semidefinite at the solution x ∗ ,
some of the individual Hessians ∇ 2 φi (·) may be indefinite. One way to overcome this obstacle
is to apply the SR1 update to each of the element Hessians. This approach has proved effective
in the LANCELOT package [72], which is designed to take full advantage of partial separability.
The main limitations of this quasi-Newton approach are the cost of the step computation (7.42), which is comparable to the cost of a Newton step, and the difficulty of identifying
the partially separable structure of a function. The performance of quasi-Newton methods
is satisfactory provided that we find the finest partially separable decomposition of the
problem; see [72]. Furthermore, even when the partially separable structure is known, it
may be more efficient to compute a Newton step. For example, the modeling language AMPL
automatically detects the partially separable structure of a function f and uses it to compute
the Hessian ∇ 2 f (x).
7.5
PERSPECTIVES AND SOFTWARE
Newton–CG methods have been used successfully to solve large problems in a variety of applications. Many of these implementations are developed by engineers and
189
190
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
scientists and use problem-specific preconditioners. Freely available packages include
TN/TNBC [220] and TNPACK [275]. Software for more general problems, such as LANCELOT
[72], KNITRO/CG [50], and TRON [192], employ Newton–CG methods when applied to
unconstrained problems. Other packages, such as LOQO [294] implement Newton methods with a sparse factorization modified to ensure positive definiteness. GLTR [145] offers
a Newton–Lanczos method. There is insufficient experience to date to say whether the
Newton–Lanczos method is significantly better in practice than the Steihaug strategy given in
Algorithm 7.2.
Software for computing incomplete Cholesky preconditioners includes the ICFS [193]
and MA57 [166] packages. A preconditioner for Newton–CG based on limited-memory
BFGS approximations is provided in PREQN [209].
Limited-memory BFGS methods are implemented in LBFGS [194] and M1QN3 [122];
see Gill and Leonard [125] for a variant that requires less storage and appears to be quite
efficient. The compact limited-memory representations of Section 7.2 are used in LBFGS-B
[322], IPOPT [301], and KNITRO.
The LANCELOT package exploits partial separability. It provides SR1 and BFGS quasiNewton options as well as a Newton methods. The step computation is obtained by a
preconditioned conjugate gradient iteration using trust regions. If f is partially separable, a
general affine transformation will not in general preserve the partially separable structure.
The quasi-Newton method for partially separable functions described in Section 7.4 is not
invariant to affine transformations of the variables, but this is not a drawback because the
method is invariant under transformations that preserve separability.
NOTES AND REFERENCES
A complete study of inexact Newton methods is given in [74]. For a discussion
of the Newton–Lanczos method see [145]. Other iterative methods for the solution of a
trust-region problem have been proposed by Hager [160], and by Rendl and Wolkowicz
[263].
For further discussion on the L-BFGS method see Nocedal [228], Liu and Nocedal
[194], and Gilbert and Lemaréchal [122]. The last paper also discusses various ways in
which the scaling parameter can be chosen. Algorithm 7.4, the two-loop L-BFGS recursion,
constitutes an economical procedure for computing the product Hk ∇ f k . It is based, however,
on the specific form of the BFGS update formula (7.16), and recursions of this type have not
yet been developed (and may not exist) for other members of the Broyden class (for instance,
the SR1 and DFP methods). Our discussion of compact representations of limited-memory
matrices is based on Byrd, Nocedal, and Schnabel [52].
Sparse quasi-Newton updates have been studied by Toint [288, 289] and Fletcher et
al. [102, 104], among others. The concept of partial separability was introduced by Griewank
and Toint [156, 155]. For an extensive treatment of the subject see Conn, Gould, and Toint
[72].
7.5.
PERSPECTIVES
AND
SOFTWARE
✐
EXERCISES
✐
7.1 Code Algorithm 7.5, and test it on the extended Rosenbrock function
f (x) n/2
2
α(x2i − x2i−1
)2 + (1 − x2i−1 )2 ,
i1
where α is a parameter that you can vary (for example, 1 or 100). The solution is x ∗ (1, 1, . . . , 1)T , f ∗ 0. Choose the starting point as (−1, −1, . . . , −1)T . Observe the
behavior of your program for various values of the memory parameter m.
✐
7.2 Show that the matrix Ĥk+1 in (7.21) is singular.
✐
7.3 Derive the formula (7.23) under the assumption that line searches are exact.
✐
7.4 Consider limited-memory SR1 updating based on (7.30). Explain how the storage
can be cut in half if the basic matrix Bk0 is kept fixed for all k. (Hint: Consider the matrix
Q k [q0 , . . . , qk−1 ] Yk − B0 Sk .)
✐
7.5 Write the function defined by
f (x) x2 x3 e x1 +x3 −x4 + (x2 x3 )2 + (x3 − x4 )
in the form (7.40). In particular, give the definition of each of the compactifying
transformations Ui .
✐
7.6 Does the approximation B obtained by the partially separable quasi-Newton
updating (7.38), (7.41) satisfy the secant equation Bs y?
✐
7.7 The minimum surface problem is a classical application of the calculus of variations and can be found in many textbooks. We wish to find the surface of minimum
area, defined on the unit square, that interpolates a prescribed continuous function on the
boundary of the square. In the standard discretization of this problem, the unknowns are
the values of the sought-after function z(x, y) on a q × q rectangular mesh of points over
the unit square.
More specifically, we divide each edge of the square into q intervals of equal length,
yielding (q + 1)2 grid points. We label the grid points as
x(i−1)(q+1)+1 , . . . , xi(q+1)
for i 1, 2, ..., q + 1,
so that each value of i generates a line. With each point we associate a variable z i that
represents the height of the surface at this point. For the 4q grid points on the boundary
of the unit square, the values of these variables are determined by the given function. The
191
192
CHAPTER 7.
LARGE-SCALE UNCONSTRAINED OPTIMIZATION
optimization problem is to determine the other (q + 1)2 − 4q variables z i so that the total
surface area is minimized.
A typical subsquare in this partition looks as follows:
x j+q+1
x j+q+2
xj
x j+1
We denote this square by A j and note that its area is q 2 . The desired function is z(x, y), and
we wish to compute its surface over A j . Calculus books show that the area of the surface is
given by
+
f j (x) ≡
1+
(x,y)∈A j
∂z
∂x
2
+
∂z
∂y
2
d x d y.
Approximate the derivatives by finite differences, and show that f j has the form
1
2
q2
1
2
2
f j (x) 2 1 + [(x j − x j+q+1 ) + (x j+1 − x j+q ) ] .
q
2
✐
(7.43)
7.8 Compute the gradient of the element function (7.43) with respect to the full
vector x. Show that it contains at most four nonzeros, and that two of these four nonzero
components are negatives of the other two. Compute the Hessian of f j , and show that,
among the 16 nonzeros, only three different magnitudes are represented. Also show that
this Hessian is singular.
This is page 193
Printer: Opaque this
CHAPTER
8
Calculating
Derivatives
Most algorithms for nonlinear optimization and nonlinear equations require knowledge of
derivatives. Sometimes the derivatives are easy to calculate by hand, and it is reasonable
to expect the user to provide code to compute them. In other cases, the functions are too
complicated, so we look for ways to calculate or approximate the derivatives automatically.
A number of interesting approaches are available, of which the most important are probably
the following.
Finite Differencing. This technique has its roots in Taylor’s theorem (see Chapter 2). By
observing the change in function values in response to small perturbations of the unknowns
194
CHAPTER 8.
CALCULATING DERIVATIVES
near a given point x, we can estimate the response to infintesimal perturbations, that is,
the derivatives. For instance, the partial derivative of a smooth function f : IRn → IR with
respect to the ith variable xi can be approximated by the central-difference formula
∂f
f (x + ei ) − f (x − ei )
,
≈
∂ xi
2
where is a small positive scalar and ei is the ith unit vector, that is, the vector whose
elements are all 0 except for a 1 in the ith position.
Automatic Differentiation. This technique takes the view that the computer code for
evaluating the function can be broken down into a composition of elementary arithmetic
operations, to which the chain rule (one of the basic rules of calculus) can be applied. Some
software tools for automatic differentiation (such as ADIFOR [25]) produce new code that
calculates both function and derivative values. Other tools (such as ADOL-C [154]) keep a
record of the elementary computations that take place while the function evaluation code
for a given point x is executing on the computer. This information is processed to produce
the derivatives at the same point x.
Symbolic Differentiation. In this technique, the algebraic specification for the function f is
manipulated by symbolic manipulation tools to produce new algebraic expressions for each
component of the gradient. Commonly used symbolic manipulation tools can be found in
the packages Mathematica [311], Maple [304], and Macsyma [197].
In this chapter we discuss the first two approaches: finite differencing and automatic
differentiation.
The usefulness of derivatives is not restricted to algorithms for optimization. Modelers
in areas such as design optimization and economics are often interested in performing
post-optimal sensitivity analysis, in which they determine the sensitivity of the optimum to
small perturbations in the parameter or constraint values. Derivatives are also important in
other areas such as nonlinear differential equations and simulation.
8.1
FINITE-DIFFERENCE DERIVATIVE APPROXIMATIONS
Finite differencing is an approach to the calculation of approximate derivatives whose
motivation (like that of so many algorithms in optimization) comes from Taylor’s theorem.
Many software packages perform automatic calculation of finite differences whenever the
user is unable or unwilling to supply code to calculate exact derivatives. Although they yield
only approximate values for the derivatives, the results are adequate in many situations.
By definition, derivatives are a measure of the sensitivity of the function to infinitesimal
changes in the values of the variables. Our approach in this section is to make small, finite
perturbations in the values of x and examine the resulting differences in the function values.
8.1.
FINITE-DIFFERENCE DERIVATIVE APPROXIMATIONS
By taking ratios of the function difference to variable difference, we obtain approximations
to the derivatives.
APPROXIMATING THE GRADIENT
An approximation to the gradient vector ∇ f (x) can be obtained by evaluating the
function f at (n + 1) points and performing some elementary arithmetic. We describe this
technique, along with a more accurate variant that requires additional function evaluations.
A popular formula for approximating the partial derivative ∂ f /∂ xi at a given point x
is the forward-difference, or one-sided-difference, approximation, defined as
∂f
f (x + ei ) − f (x)
(x) ≈
.
∂ xi
(8.1)
The gradient can be built up by simply applying this formula for i 1, 2, . . . , n. This
process requires evaluation of f at the point x as well as the n perturbed points x + ei ,
i 1, 2, . . . , n: a total of (n + 1) points.
The basis for the formula (8.1) is Taylor’s theorem, Theorem 2.1 in Chapter 2. When
f is twice continuously differentiable, we have
f (x + p) f (x) + ∇ f (x)T p + 12 p T ∇ 2 f (x + t p) p,
some t ∈ (0, 1)
(8.2)
(see (2.6)). If we choose L to be a bound on the size of ∇ 2 f (·) in the region of interest,
it follows directly from this formula that the last term in this expression is bounded by
(L/2) p2 , so that
f (x + p) − f (x) − ∇ f (x)T p ≤ (L/2) p2 .
(8.3)
We now choose the vector p to be ei , so that it represents a small change in the value
of a single component of x (the ith component). For this p, we have that ∇ f (x)T p ∇ f (x)T ei ∂ f /∂ xi , so by rearranging (8.3), we conclude that
f (x + ei ) − f (x)
∂f
+δ ,
(x) ∂ xi
where |δ | ≤ (L/2) .
(8.4)
We derive the forward-difference formula (8.1) by simply ignoring the error term δ in this
expression, which becomes smaller and smaller as approaches zero.
An important issue in implementing the formula (8.1) is the choice of the parameter
. The error expression (8.4) suggests that we should choose as small as possible. Unfortunately, this expression ignores the roundoff errors that are introduced when the function
f is evaluated on a real computer, in floating-point arithmetic. From our discussion in the
Appendix (see (A.30) and (A.31)), we know that the quantity u known as unit roundoff
195
196
CHAPTER 8.
CALCULATING DERIVATIVES
is crucial: It is a bound on the relative error that is introduced whenever an arithmetic
operation is performed on two floating-point numbers. (u is about 1.1 × 10−16 in doubleprecision IEEE floating-point arithmetic.) The effect of these errors on the final computed
value of f depends on the way in which f is computed. It could come from an arithmetic
formula, or from a differential equation solver, with or without refinement.
As a rough estimate, let us assume simply that the relative error in the computed f is
bounded by u, so that the computed values of f (x) and f (x + ei ) are related to the exact
values in the following way:
|comp( f (x)) − f (x)| ≤ uL f ,
|comp( f (x + ei )) − f (x + ei )| ≤ uL f ,
where comp(·) denotes the computed value, and L f is a bound on the value of | f (·)| in the
region of interest. If we use these computed values of f in place of the exact values in (8.4)
and (8.1), we obtain an error that is bounded by
(L/2) + 2uL f / .
(8.5)
Naturally, we would like to choose to make this error as small as possible; it is easy to see
that the minimizing value is
2
4L f u
.
L
If we assume that the problem is well scaled, then the ratio L f /L (the ratio of function
values to second derivative values) does not exceed a modest size. We can conclude that the
following choice of is fairly close to optimal:
√
u.
(8.6)
(In fact, this value is used in many of the optimization software packages that use finite
differencing as an option for estimating derivatives.) For this value of , we have from (8.5)
√
that the total error in the forward-difference approximation is fairly close to u.
A more accurate approximation to the derivative can be obtained by using the centraldifference formula, defined as
∂f
f (x + ei ) − f (x − ei )
.
(x) ≈
∂ xi
2
(8.7)
As we show below, this approximation is more accurate than the forward-difference approximation (8.1). It is also about twice as expensive, since we need to evaluate f at the points
x and x ± ei , i 1, 2, . . . , n: a total of 2n + 1 points.
8.1.
FINITE-DIFFERENCE DERIVATIVE APPROXIMATIONS
The basis for the central difference approximation is again Taylor’s theorem. When
the second derivatives of f exist and are Lipschitz continuous, we have from (8.2) that
f (x + p) f (x) + ∇ f (x)T p + 12 p T ∇ 2 f (x + t p) p
for some t ∈ (0, 1)
f (x) + ∇ f (x) p + p ∇ f (x) p + O p3 .
T
1
2
T
2
(8.8)
By setting p ei and p − ei , respectively, we obtain
f (x + ei ) f (x) +
∂f
1
+
∂ xi
2
2∂
f (x − ei ) f (x) −
∂f
1
+
∂ xi
2
2∂
2
f
∂ xi2
2
f
∂ xi2
+O
+O
3
3
,
.
(Note that the final error terms in these two expressions are generally not the same, but they
are both bounded by some multiple of 3 .) By subtracting the second equation from the
first and dividing by 2 , we obtain the expression
f (x + ei ) − f (x − ei )
∂f
+O 2 .
(x) ∂ xi
2
We see from this expression that the error is O 2 , as compared to the O( ) error in the
forward-difference formula (8.1). However, when we take evaluation error in f into account,
the accuracy that can be achieved in practice is less impressive; the same assumptions that
were used to derive (8.6) lead to an optimal choice of of about u1/3 and an error of about
u2/3 . In some situations, the extra few digits of accuracy may improve the performance of
the algorithm enough to make the extra expense worthwhile.
APPROXIMATING A SPARSE JACOBIAN
Consider now the case of a vector function r : IRn → IRm , such as the residual vector
that we consider in Chapter 10 or the system of nonlinear equations from Chapter 11. The
matrix J (x) of first derivatives for this function is defined as follows:
⎡
∂r j
J (x) ∂ xi
j1,2,...,m
i1,2,...,n
∇r1 (x)T
⎢
⎢ ∇r2 (x)T
⎢
⎢
..
⎢
.
⎣
∇rm (x)T
⎤
⎥
⎥
⎥
⎥,
⎥
⎦
(8.9)
where r j , j 1, 2, . . . , m are the components of r . The techniques described in the previous
197
198
CHAPTER 8.
CALCULATING DERIVATIVES
section can be used to evaluate the full Jacobian J (x) one column at a time. When r is twice
continuously differentiable, we can use Taylor’s theorem to deduce that
r (x + p) − r (x) − J (x) p ≤ (L/2) p2 ,
(8.10)
where L is a Lipschitz constant for J in the region of interest. If we require an approximation
to the Jacobian–vector product J (x) p for a given vector p (as is the case with inexact Newton
methods for nonlinear systems of equations; see Section 11.1), this expression immediately
suggests choosing a small nonzero and setting
J (x) p ≈
r (x + p) − r (x)
,
(8.11)
an approximation that is accurate to O( ). A two-sided approximation can be derived from
the formula (8.7).
If an approximation to the full Jacobian J (x) is required, we can compute it a column
at a time, analogously to (8.1), by setting set p ei in (8.10) to derive the following
estimate of the ith column:
∂r
r (x + ei ) − r (x)
(x) ≈
.
∂ xi
(8.12)
A full Jacobian estimate can be obtained at a cost of n + 1 evaluations of the function r .
When the Jacobian is sparse, however, we can often obtain the estimate at a much lower cost,
sometimes just three or four evaluations of r . The key is to estimate a number of different
columns of the Jacobian simultaneously, by judicious choices of the perturbation vector p
in (8.10).
We illustrate the technique with a simple example. Consider the function r : IRn → IRn
defined by
⎡
2(x23 − x12 )
⎢
⎢ 3(x23 − x12 ) + 2(x33 − x22 )
⎢
⎢
3(x33 − x22 ) + 2(x43 − x32 )
r (x) ⎢
⎢
⎢
..
⎢
.
⎣
3
2
3(xn − xn−1
)
⎤
⎥
⎥
⎥
⎥
⎥.
⎥
⎥
⎥
⎦
(8.13)
Each component of r depends on just two or three components of x, so that each row of the
Jacobian contains only two or three nonzero elements. For the case of n 6, the Jacobian
8.1.
FINITE-DIFFERENCE DERIVATIVE APPROXIMATIONS
has the following structure:
⎡
×
⎢
⎢ ×
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
⎤
×
×
×
×
×
×
×
×
×
×
×
⎥
⎥
⎥
⎥
⎥
⎥,
⎥
⎥
⎥
× ⎦
×
×
(8.14)
where each cross represents a nonzero element, with zeros represented by a blank space.
Staying for the moment with the case n 6, suppose that we wish to compute a finitedifference approximation to the Jacobian. (Of course, it is easy to calculate this particular
Jacobian by hand, but there are complicated functions with similar structure for which
hand calculation is more difficult.) A perturbation p e1 to the first component of x
will affect only the first and second components of r . The remaining components will be
unchanged, so that the right-hand-side of formula (8.12) will correctly evaluate to zero in
the components 3, 4, 5, 6. It is wasteful, however, to reevaluate these components of r when
we know in advance that their values are not affected by the perturbation. Instead, we look
for a way to modify the perturbation vector so that it does not have any further effect on
components 1 and 2, but does produce a change in some of the components 3, 4, 5, 6, which
we can then use as the basis of a finite-difference estimate for some other column of the
Jacobian. It is not hard to see that the additional perturbation e4 has the desired property:
It alters the 3rd, 4th, and 5th elements of r , but leaves the 1st and 2nd elements unchanged.
The changes in r as a result of the perturbations e1 and e4 do not interfere with each
other.
To express this discussion in mathematical terms, we set
p (e1 + e4 ),
and note that
r (x + p)1,2 r (x + (e1 + e4 ))1,2 r (x + e1 )1,2
(8.15)
(where the notation [·]1,2 denotes the subvector consisting of the first and second elements),
while
r (x + p)3,4,5 r (x + (e1 + e4 ))3,4,5 r (x + e4 )3,4,5 .
By substituting (8.15) into (8.10), we obtain
r (x + p)1,2 r (x)1,2 + [J (x)e1 ]1,2 + O( 2 ).
(8.16)
199
200
CHAPTER 8.
CALCULATING DERIVATIVES
By rearranging this expression, we obtain the following difference formula for estimating
the (1, 1) and (2, 1) elements of the Jacobian matrix:
⎤
∂r1
⎢ ∂ x1 (x) ⎥
⎥ [J (x)e1 ]1,2 ≈ r (x + p)1,2 − r (x)1,2 .
⎢
⎦
⎣ ∂r2
(x)
∂ x1
⎡
(8.17)
A similar argument shows that the nonzero elements of the fourth column of the Jacobian
can be estimated by substituting (8.16) into (8.10); we obtain
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣
∂r4
(x)
∂ x3
∂r4
(x)
∂ x4
∂r4
(x)
∂ x5
⎤
⎥
⎥
⎥
⎥ [J (x)e4 ]3,4,5 ≈ r (x + p)3,4,5 − r (x)3,4,5 .
⎥
⎥
⎦
(8.18)
To summarize: We have been able to estimate two columns of the Jacobian J (x) by evaluating
the function r at the single extra point x + (e1 + e4 ).
We can approximate the remainder of J (x) in an economical manner as well. Columns
2 and 5 can be approximated by choosing p (e2 + e5 ), while we can use p (e3 + e6 )
to approximate columns 3 and 6. In total, we need 3 evaluations of the function r (after the
initial evaluation at x) to estimate the entire Jacobian matrix.
In fact, for any choice of n in (8.13) (no matter how large), three extra evaluations of r
are sufficient to approximate the entire Jacobian. The corresponding choices of perturbation
vectors p are
p (e1 + e4 + e7 + e10 + · · ·),
p (e2 + e5 + e8 + e11 + · · ·),
p (e3 + e6 + e9 + e12 + · · ·).
In the first of these vectors, the nonzero components are chosen so that no two of the
columns 1, 4, 7, . . . have a nonzero element in the same row. The same property holds for
the other two vectors and, in fact, points the way to the criterion that we can apply to general
problems to decide on a valid set of perturbation vectors.
Algorithms for choosing the perturbation vectors can be expressed conveniently in the
language of graphs and graph coloring. For any function r : IRn → IRm , we can construct
a column incidence graph G with n nodes by drawing an arc between nodes i and k if
there is some component of r that depends on both xi and xk . In other words, the ith and
kth columns of the Jacobian J (x) each have a nonzero element in some row j, for some
j 1, 2, . . . , m and some value of x. (The intersection graph for the function defined in
8.1.
FINITE-DIFFERENCE DERIVATIVE APPROXIMATIONS
1
2
3
4
5
6
Figure 8.1
Column incidence graph for r (x) defined in (8.13).
(8.13), with n 6, is shown in Figure 8.1.) We now assign each node a “color” according
to the following rule: Two nodes can have the same color if there is no arc that connects
them. Finally, we choose one perturbation vector corresponding to each color: If nodes
i 1 , i 2 , . . . , i have the same color, the corresponding p is (ei1 + ei2 + · · · + ei ).
Usually, there are many ways to assign colors to the n nodes in the graph in a way that
satisfies the required condition. The simplest way is just to assign each node a different color,
but since that scheme produces n perturbation vectors, it is usually not the most efficient
approach. It is generally very difficult to find the coloring scheme that uses the fewest
possible colors, but there are simple algorithms that do a good job of finding a near-optimal
coloring at low cost. Curtis, Powell, and Reid [83] and Coleman and Moré [68] provide
descriptions of some methods and performance comparisons. Newsam and Ramsdell [227]
show that by considering a more general class of perturbation vectors p, it is possible
to evaluate the full Jacobian using no more than n z evaluations of r (in addition to the
evaluation at the point x), where n z is the maximum number of nonzeros in each row
of J (x).
For some functions r with well-studied structures (those that arise from discretizations
of differential operators, or those that give rise to banded Jacobians, as in the example above),
optimal coloring schemes are known. For the tridiagonal Jacobian of (8.14) and its associated
graph in Figure 8.1, the scheme with three colors is optimal.
APPROXIMATING THE HESSIAN
In some situations, the user may be able to provide a routine to calculate the gradient
∇ f (x) but not the Hessian ∇ 2 f (x). We can obtain the Hessian by applying the techniques
described above for the vector function r to the gradient ∇ f . By using the graph coloring
techniques discussed above, sparse Hessians often can be approximated in this manner by
using considerably fewer than n perturbation vectors. This approach ignores symmetry
of the Hessian, and will usually produce a nonsymmetric approximation. We can recover
201
202
CHAPTER 8.
CALCULATING DERIVATIVES
symmetry by adding the approximation to its transpose and dividing the result by 2.
Alternative differencing approaches that take symmetry of ∇ 2 f (x) explicitly into account
are discussed below.
Some important algorithms—most notably the Newton–CG methods described in
Chapter 7—do not require knowledge of the full Hessian. Instead, each iteration requires
us to supply the Hessian–vector product ∇ 2 f (x) p, for a given vector p. We can obtain an
approximation to this matrix-vector product by appealing once again to Taylor’s theorem.
When second derivatives of f exist and are Lipschitz continuous near x, we have
∇ f (x + p) ∇ f (x) + ∇ 2 f (x) p + O( 2 ),
(8.19)
so that
∇ 2 f (x) p ≈
∇ f (x + p) − ∇ f (x)
(8.20)
(see also (7.10)). The approximation error is O( ), and the cost of obtaining the approximation is a single gradient evaluation at the point x + p. The formula (8.20) corresponds
to the forward-difference approximation (8.1). A central-difference formula like (8.7) can
be derived by evaluating ∇ f (x − p) as well.
For the case in which even gradients are not available, we can use Taylor’s theorem
once again to derive formulae for approximating the Hessian that use only function values.
The main tool is the formula (8.8): By substituting the vectors p ei , p e j , and
p (ei + e j ) into this formula and combining the results appropriately, we obtain
f (x + ei + e j ) − f (x + ei ) − f (x + e j ) + f (x)
∂2 f
(x) + O( ).
2
∂ xi ∂ x j
(8.21)
If we wished to approximate every element of the Hessian with this formula, then we would
need to evaluate f at x + (ei + e j ) for all possible i and j (a total of n(n + 1)/2 points)
as well as at the n points x + ei , i 1, 2, . . . , n. If the Hessian is sparse, we can, of course,
reduce this operation count by skipping the evaluation whenever we know the element
∂ 2 f /∂ xi ∂ x j to be zero.
APPROXIMATING A SPARSE HESSIAN
We noted above that a Hessian approximation can be obtained by applying finitedifference Jacobian estimation techniques to the gradient ∇ f , treated as a vector function.
We now show how symmetry of the Hessian ∇ 2 f can be used to reduce the number of
perturbation vectors p needed to obtain a complete approximation, when the Hessian
is sparse. The key observation is that, because of symmetry, any estimate of the element
[∇ 2 f (x)]i, j ∂ 2 f (x)/∂ xi ∂ x j is also an estimate of its symmetric counterpart [∇ 2 f (x)] j,i .
8.1.
FINITE-DIFFERENCE DERIVATIVE APPROXIMATIONS
We illustrate the point with the simple function f : IRn → IR defined by
f (x) x1
n
i 2 xi2 .
(8.22)
i1
It is easy to show that the Hessian ∇ 2 f has the “arrowhead” structure depicted below, for
the case of n 6:
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
⎤
⎥
⎥
⎥
⎥
⎥
⎥.
⎥
⎥
⎥
⎦
(8.23)
×
If we were to construct the intersection graph for the function ∇ f (analogous to Figure 8.1),
we would find that every node is connected to every other node, for the simple reason that
row 1 has a nonzero in every column. According to the rule for coloring the graph, then, we
would have to assign a different color to every node, which implies that we would need to
evaluate ∇ f at the n + 1 points x and x + ei for i 1, 2, . . . , n.
We can construct a much more efficient scheme by taking the symmetry into account.
Suppose we first use the perturbation vector p e1 to estimate the first column of ∇ 2 f (x).
Because of symmetry, the same estimates apply to the first row of ∇ 2 f . From (8.23), we see
that all that remains is to find the diagonal elements ∇ 2 f (x)22 , ∇ 2 f (x)33 , . . . , ∇ 2 f (x)66 .
The intersection graph for these remaining elements is completely disconnected, so we can
assign them all the same color and choose the corresponding perturbation vector to be
p (e2 + e3 + · · · + e6 ) (0, 1, 1, 1, 1, 1)T .
(8.24)
Note that the second component of ∇ f is not affected by the perturbations in components
3, 4, 5, 6 of the unknown vector, while the third component of ∇ f is not affected by
perturbations in components 2, 4, 5, 6 of x, and so on. As in (8.15) and (8.16), we have for
each component i that
∇ f (x + p)i ∇ f (x + (e2 + e3 + · · · + e6 ))i ∇ f (x + ei )i .
By applying the forward-difference formula (8.1) to each of these individual components,
we then obtain
∂2 f
∇ f (x + ei )i − ∇ f (x)i
∇ f (x + p)i − ∇ f (x)i
(x) ≈
,
∂ xi2
i 2, 3, . . . , 6.
203
204
CHAPTER 8.
CALCULATING DERIVATIVES
By exploiting symmetry, we have been able to estimate the entire Hessian by evaluating ∇ f
only at x and two other points.
Again, graph-coloring techniques can be used to choose the perturbation vectors p
economically. We use the adjacency graph in place of the intersection graph described earlier.
The adjacency graph has n nodes, with arcs connecting nodes i and k whenever i k and
∂ 2 f (x)/(∂ xi ∂ xk ) 0 for some x. The requirements on the coloring scheme are a little more
complicated than before, however. We require not only that connected nodes have different
colors, but also that any path of length 3 through the graph contain at least three colors. In
other words, if there exist nodes i 1 , i 2 , i 3 , i 4 in the graph that are connected by arcs (i 1 , i 2 ),
(i 2 , i 3 ), and (i 3 , i 4 ), then at least three different colors must be used in coloring these four
nodes. See Coleman and Moré [69] for an explanation of this rule and for algorithms to
compute valid colorings. The perturbation vectors are constructed as before: Whenever the
nodes i 1 , i 2 , . . . , i have the same color, we set the corresponding perturbation vector to be
p (ei1 + ei2 + · · · + ei ).
8.2
AUTOMATIC DIFFERENTIATION
Automatic differentiation is the generic name for techniques that use the computational
representation of a function to produce analytic values for the derivatives. Some techniques
produce code for the derivatives at a general point x by manipulating the function code
directly. Other techniques keep a record of the computations made during the evaluation
of the function at a specific point x and then review this information to produce a set of
derivatives at x.
Automatic differentiation techniques are founded on the observation that any function, no matter how complicated, is evaluated by performing a sequence of simple elementary
operations involving just one or two arguments at a time. Two-argument operations include
addition, multiplication, division, and the power operation a b . Examples of single-argument
operations include the trigonometric, exponential, and logarithmic functions. Another common ingredient of the various automatic differentiation tools is their use of the chain rule.
This is the well-known rule from elementary calculus that says that if h is a function of the
vector y ∈ IRm , which is in turn a function of the vector x ∈ IRn , we can write the derivative
of h with respect to x as follows:
∇x h(y(x)) m
∂h
∇ yi (x).
∂ yi
i1
(8.25)
See Appendix A for further details.
There are two basic modes of automatic differentiation: the forward and reverse modes.
The difference between them can be illustrated by a simple example. We work through such
8.2.
AUTOMATIC DIFFERENTIATION
an example below, and indicate how the techniques can be extended to general functions,
including vector functions.
AN EXAMPLE
Consider the following function of 3 variables:
f (x) (x1 x2 sin x3 + e x1 x2 )/x3 .
(8.26)
Figure 8.2 shows how the evaluation of this function can be broken down into its elementary
operations and also indicates the partial ordering associated with these operations. For
instance, the multiplication x1 ∗ x2 must take place prior to the exponentiation e x1 x2 , or else
we would obtain the incorrect result (e x1 )x2 . This graph introduces the intermediate variables
x4 , x5 , . . . that contain the results of intermediate computations; they are distinguished from
the independent variables x1 , x2 , x3 that appear at the left of the graph. We can express the
evaluation of f in arithmetic terms as follows:
x4 x1 ∗ x2 ,
x5 sin x3 ,
x 6 e x4 ,
(8.27)
x7 x4 ∗ x5 ,
x8 x6 + x7 ,
x9 x8 /x3 .
The final node x9 in Figure 8.2 contains the function value f (x). In the terminology
of graph theory, node i is the parent of node j, and node j the child of node i, whenever
there is a directed arc from i to j. Any node can be evaluated when the values of all its
parents are known, so computation flows through the graph from left to right. Flow of
x1
*
x4
exp
x6
x2
x3
Figure 8.2
sin
x
5
*
x7
+
x8
/
Computational graph for f (x) defined in (8.26).
x9
205
206
CHAPTER 8.
CALCULATING DERIVATIVES
computation in this direction is known as a forward sweep. It is important to emphasize
that software tools for automatic differentiation do not require the user to break down the
code for evaluating the function into its elements, as in (8.27). Identification of intermediate
quantities and construction of the computational graph is carried out, explicitly or implicitly,
by the software tool itself.
THE FORWARD MODE
In the forward mode of automatic differentiation, we evaluate and carry forward
a directional derivative of each intermediate variable xi in a given direction p ∈ IRn ,
simultaneously with the evaluation of xi itself. For the three-variable example above, we use
the following notation for the directional derivative for p associated with each variable:
def
D p xi (∇xi )T p 3
∂ xi
pj,
∂x j
j1
i 1, 2, . . . , 9,
(8.28)
where ∇ indicates the gradient with respect to the three independent variables. Our goal
is to evaluate D p x9 , which is the same as the directional derivative ∇ f (x)T p. We note
immediately that initial values D p xi for the independent variables xi , i 1, 2, 3, are simply
the components p1 , p2 , p3 of p. The direction p is referred to as the seed vector.
As soon as the value of xi at any node is known, we can find the corresponding value
of D p xi from the chain rule. For instance, suppose we know the values of x4 , D p x4 , x5 , and
D p x5 , and we are about to calculate x7 in Figure 8.2. We have that x7 x4 x5 ; that is, x7
is a function of the two variables x4 and x5 , which in turn are functions of x1 , x2 , x3 . By
applying the rule (8.25), we have that
∇x7 ∂ x7
∂ x7
∇x4 +
∇x5 x5 ∇x4 + x4 ∇x5 .
∂ x4
∂ x5
By taking the inner product of both sides of this expression with p and applying the definition
(8.28), we obtain
D p x7 ∂ x7
∂ x7
D p x4 +
D p x5 x5 D p x4 + x4 D p x5 .
∂ x4
∂ x5
(8.29)
The directional derivatives D p xi are therefore evaluated side by side with the intermediate
results xi , and at the end of the process we obtain D p x9 D p f ∇ f (x)T p.
The principle of the forward mode is straightforward enough, but what of its practical
implementation and computational requirements? First, we repeat that the user does not
need to construct the computational graph, break the computation down into elementary
operations as in (8.27), or identify intermediate variables. The automatic differentiation
software should perform these tasks implicitly and automatically. Nor is it necessary to store
8.2.
AUTOMATIC DIFFERENTIATION
the information xi and D p xi for every node of the computation graph at once (which is just
as well, since this graph can be very large for complicated functions). Once all the children
of any node have been evaluated, its associated values xi and D p xi are not needed further
and may be overwritten in storage.
The key to practical implementation is the side-by-side evaluation of xi and D p xi . The
automatic differentiation software associates a scalar D p w with any scalar w that appears
in the evaluation code. Whenever w is used in an arithmetic computation, the software
performs an associated operation (based on the chain rule) with the gradient vector D p w.
For instance, if w is combined in a division operation with another value y to produce a
new value z, that is,
z←
w
,
y
we use w, z, D p w, and D p y to evaluate the directional derivative D p z as follows:
Dpz ←
1
w
D p w − 2 D p y.
y
y
(8.30)
To obtain the complete gradient vector, we can carry out this procedure simultaneously
for the n seed vectors p e1 , e2 , . . . , en . By the definition (8.28), we see that p e j implies
that D p f ∂ f /∂ x j , j 1, 2, . . . , n. We note from the example (8.30) that the additional
cost of evaluating f and ∇ f (over the cost of evaluating f alone) may be significant. In
this example, the single division operation on w and y needed to calculate z gives rise
to approximately 2n multiplications and n additions in the computation of the gradient
elements De j z, j 1, 2, . . . , n. It is difficult to obtain an exact bound on the increase in
computation, since the costs of retrieving and storing the data should also be taken into
account. The storage requirements may also increase by a factor as large as n, since we
now have to store n additional scalars De j xi , j 1, 2, . . . , n, alongside each intermediate
variable xi . It is usually possible to make savings by observing that many of these quantities
are zero, particularly in the early stages of the computation (that is, toward the left of the
computational graph), so sparse data structures can be used to store the vectors De j xi ,
j 1, 2, . . . , n (see [27]).
The forward mode of automatic differentiation can be implemented by means of a
precompiler, which transforms function evaluation code into extended code that evaluates
the derivative vectors as well. An alternative approach is to use the operator-overloading
facilities available in languages such as C++ to transparently extend the data structures and
operations in the manner described above.
THE REVERSE MODE
The reverse mode of automatic differentiation does not perform function and gradient
evaluations concurrently. Instead, after the evaluation of f is complete, it recovers the partial
207
208
CHAPTER 8.
CALCULATING DERIVATIVES
derivatives of f with respect to each variable xi —independent and intermediate variables
alike—by performing a reverse sweep of the computational graph. At the conclusion of this
process, the gradient vector ∇ f can be assembled from the partial derivatives ∂ f /∂ xi with
respect to the independent variables xi , i 1, 2, . . . , n.
Instead of the gradient vectors D p xi used in the forward mode, the reverse mode
associates a scalar variable x̄i with each node in the graph; information about the partial
derivative ∂ f /∂ xi is accumulated in x̄i during the reverse sweep. The x̄i are sometimes
called the adjoint variables, and we initialize their values to zero, with the exception
of the rightmost node in the graph (node N , say), for which we set x̄ N 1. This
choice makes sense because x N contains the final function value f , so we have ∂ f /
∂ x N 1.
The reverse sweep makes use of the following observation, which is again based on
the chain rule (8.25): For any node i, the partial derivative ∂ f /∂ xi can be built up from
the partial derivatives ∂ f /∂ x j corresponding to its child nodes j according to the following
formula:
∂f
∂ xi
j a child of i
∂ f ∂x j
.
∂ x j ∂ xi
(8.31)
For each node i, we add the right-hand-side term in (8.31) to x̄i as soon as it becomes
known; that is, we perform the operation
x̄i + ∂ f ∂x j
.
∂ x j ∂ xi
(8.32)
(In this expression and the ones below, we use the arithmetic notation of the programming
language C, in which x+ a means x ← x + a.) Once contributions have been received
from all the child nodes of i, we have x̄i ∂ f /∂ xi , so we declare node i to be “finalized.”
At this point, node i is ready to contribute a term to the summation for each of its parent
nodes according to the formula (8.31). The process continues in this fashion until all nodes
are finalized. Note that for derivative evaluation, the flow of computation in the graph
is from children to parents—the opposite direction to the computation flow for function
evaluation.
During the reverse sweep, we work with numerical values, not with formulae or
computer code involving the variables xi or the partial derivatives ∂ f /∂ xi . During the
forward sweep—the evaluation of f —we not only calculate the values of each variable xi ,
but we also calculate and store the numerical values of each partial derivative ∂ x j /∂ xi . Each
of these partial derivatives is associated with a particular arc of the computational graph.
The numerical values of ∂ x j /∂ xi computed during the forward sweep are then used in the
formula (8.32) during the reverse sweep.
We illustrate the reverse mode for the example function (8.26). In Figure 8.3 we fill
in the graph of Figure 8.2 for a specific evaluation point x (1, 2, π/2)T , indicating the
8.2.
1
AUTOMATIC DIFFERENTIATION
p(4,1)=2
*
2
p(4,2)=1
π /2
sin
p(5,3)=0
2
exp
p(6,4)=e
e2
p(8,6)=1
p(7,4)=1
1
*
p(7,5)=2
2
+
p(8,7)=1
2+e 2
p(9,8)=2/π
/
4+2e2/ π
p(9,3)=(84e 2)/π 2
Figure 8.3 Computational graph for f (x) defined in (8.26) showing numerical
values of intermediate values and partial derivatives for the point x (1, 2, π/2)T .
Notation: p( j, i) ∂ x j /∂ xi .
numerical values of the intermediate variables x4 , x5 , . . . , x9 associated with each node and
the partial derivatives ∂ x j /∂ xi associated with each arc.
As mentioned above, we initialize the reverse sweep by setting all the adjoint variables
x̄i to zero, except for the rightmost node, for which we have x̄9 1. Since f (x) x9 and
since node 9 has no children, we have x̄9 ∂ f /∂ x9 , and so we can immediately declare node
9 to be finalized.
Node 9 is the child of nodes 3 and 8, so we use formula (8.32) to update the values of
x̄3 and x̄8 as follows:
∂ f ∂ x9
2 + e2
−8 − 4e2
−
,
∂ x9 ∂ x3
(π/2)2
π2
2
∂ f ∂ x9
1
.
x̄8 + ∂ x9 ∂ x8
π/2
π
x̄3 + (8.33a)
(8.33b)
Node 3 is not finalized after this operation; it still awaits a contribution from its other child,
node 5. On the other hand, node 9 is the only child of node 8, so we can declare node 8 to
be finalized with the value ∂∂xf8 2/π. We can now update the values of x̄i at the two parent
nodes of node 8 by applying the formula (8.32) once again; that is,
∂ f ∂ x8
2
;
∂ x8 ∂ x6
π
∂ f ∂ x8
2
x̄7 + .
∂ x8 ∂ x7
π
x̄6 + At this point, nodes 6 and 7 are finalized, so we can use them to update nodes 4 and 5. At
209
210
CHAPTER 8.
CALCULATING DERIVATIVES
the end of this process, when all nodes are finalized, nodes 1, 2, and 3 contain
⎡
x̄1
⎤
⎡
⎢
⎥
⎢
⎣ x̄2 ⎦ ∇ f (x) ⎣
x̄3
(4 + 4e2 )/π
(2 + 2e2 )/π
⎤
⎥
⎦,
(−8 − 4e2 )/π 2
and the derivative computation is complete.
The main appeal of the reverse mode is that its computational complexity is low for
the scalar functions f : IRn → IR discussed here. The extra arithmetic associated with
the gradient computation is at most four or five times the arithmetic needed to evaluate
the function alone. Taking the division operation in (8.33) as an example, we see that two
multiplications, a division, and an addition are required for (8.33a), while a division and
an addition are required for (8.33b). This is about five times as much work as the single
division involving these nodes that was performed during the forward sweep.
As we noted above, the forward mode may require up to n times more arithmetic
to compute the gradient ∇ f than to compute the function f alone, making it appear
uncompetitive with the reverse mode. When we consider vector functions r : IRn → IRm ,
the relative costs of the forward and reverse modes become more similar as m increases, as
we describe in the next section.
An apparent drawback of the reverse mode is the need to store the entire computational
graph, which is needed for the reverse sweep. In principle, storage of this graph is not too difficult to implement. Whenever an elementary operation is performed, we can form and store
a new node containing the intermediate result, pointers to the (one or two) parent nodes, and
the partial derivatives associated with these arcs. During the reverse sweep, the nodes can be
read in the reverse order to that in which they were written, giving a particularly simple access
pattern. The process of forming and writing the graph can be implemented as a straightforward extension to the elementary operations via operator overloading (as in ADOL-C [154]).
The reverse sweep/gradient evaluation can be invoked as a simple function call.
Unfortunately, the computational graph may require a huge amount of storage. If each
node can be stored in 20 bytes, then a function that requires one second of evaluation time
on a 100 megaflop computer may produce a graph of up to 2 gigabytes in size. The storage
requirements can be reduced, at the cost of some extra arithmetic, by performing partial
forward and reverse sweeps on pieces of the computational graph, reevaluating portions of
the graph as needed rather than storing the whole structure. Descriptions of this approach,
sometimes known as checkpointing, can be found in Griewank [150] and Grimm, Pottier, and
Rostaing-Schmidt [157]. An implementation of checkpointing in the context of variational
data assimilation can be found in Restrepo, Leaf, and Griewank [264] .
VECTOR FUNCTIONS AND PARTIAL SEPARABILITY
So far, we have looked at automatic differentiation of general scalar-valued functions
f : IRn → IR. In nonlinear least-squares problems (Chapter 10) and nonlinear equations
8.2.
AUTOMATIC DIFFERENTIATION
(Chapter 11), we have to deal with vector functions r : IRn → IRm with m components
r j , j 1, 2, . . . , m. The rightmost column of the computational graph then consists of m
nodes, none of which has any children, in place of the single node described above. The
forward and reverse modes can be adapted in straightforward ways to find the Jacobian
J (x), the m × n matrix defined in (8.9).
Besides their applications to least-squares and nonlinear-equations problems, automatic differentiation of vector functions is a useful technique for dealing with partially
separable functions. We recall that partial separability is commonly observed in large-scale
optimization, and we saw in Chapter 7 that there exist efficient quasi-Newton procedures for
the minimization of objective functions with this property. Since an automatic procedure for
detecting the decomposition of a given function f into its partially separable representation
was developed recently by Gay [118], it has become possible to exploit the efficiencies that
accrue from this property without asking much information from the user.
In the simplest sense, a function f is partially separable if we can express it in the form
f (x) ne
f i (x),
(8.34)
i1
where each element function f i (·) depends on just a few components of x. If we construct
the vector function r from the partially separable components, that is,
⎡
⎢
⎢
⎢
r (x) ⎢
⎢
⎣
f 1 (x)
f 2 (x)
..
.
⎤
⎥
⎥
⎥
⎥,
⎥
⎦
f ne (x)
it follows from (8.34) that
∇ f (x) J (x)T e,
(8.35)
where, as usual, e (1, 1, . . . , 1)T . Because of the partial separability property, most
columns of J (x) contain just a few nonzeros. This structure makes it possible to calculate
J (x) efficiently by applying graph-coloring techniques, as we discuss below. The gradient
∇ f (x) can then be recovered from the formula (8.35).
In constrained optimization, it is often beneficial to evaluate the objective function f
and the constraint functions ci , i ∈ I ∪E, simultaneously. By doing so, we can take advantage
of common expressions (which show up as shared intermediate nodes in the computation
graph) and thus can reduce the total workload. In this case, the vector function r can be
211
212
CHAPTER 8.
CALCULATING DERIVATIVES
defined as
f (x)
r (x) c j (x)
.
j∈I∪E
An example of shared intermediate nodes was seen in Figure 8.2, where x4 is shared during
the computation of x6 and x7 .
CALCULATING JACOBIANS OF VECTOR FUNCTIONS
The forward mode is the same for vector functions as for scalar functions. Given a
seed vector p, we continue to associate quantities D p xi with the node that calculates each
intermediate variable xi . At each of the rightmost nodes (containing r j , j 1, 2, . . . , m),
this variable contains the quantity D p r j (∇r j )T p, j 1, 2, . . . , m. By assembling
these m quantities, we obtain J (x) p, the product of the Jacobian and our chosen vector
p. As in the case of scalar functions (m 1), we can evaluate the complete Jacobian
by setting p e1 , e2 , . . . , en and evaluating the n quantities De j xi simultaneously. For
sparse Jacobians, we can use the coloring techniques outlined above in the context of finitedifference methods to make more intelligent and economical choices of the seed vectors p.
The factor of increase in cost of arithmetic, when compared to a single evaluation of r , is
about equal to the number of seed vectors used.
The key to applying the reverse mode to a vector function r (x) is to choose seed
vectors q ∈ IRm and apply the reverse mode to the scalar functions r (x)T q. The result of
this process is the vector
⎡
∇[r (x)T q] ∇ ⎣
m
⎤
q j r j (x)⎦ J (x)T q.
j1
Instead of the Jacobian–vector product that we obtain with the forward mode, the reverse
mode yields a Jacobian-transpose–vector product. The technique can be implemented by
seeding the variables x̄i in the m dependent nodes that contain r1 , r2 , . . . , rm , with the
components q1 , q2 , . . . , qm of the vector q. At the end of the reverse sweep, the node for
independent variables x1 , x2 , . . . , xn will contain
d
r (x)T q ,
d xi
i 1, 2, . . . , n,
which are simply the components of J (x)T q.
As usual, we can obtain the full Jacobian by carrying out the process above for the m
unit vectors q e1 , e2 , . . . , em . Alternatively, for sparse Jacobians, we can apply the usual
coloring techniques to find a smaller number of seed vectors q—the only difference being
8.2.
AUTOMATIC DIFFERENTIATION
that the graphs and coloring strategies are defined with reference to the transpose J (x)T
rather than to J (x) itself. The factor of increase in the number of arithmetic operations
required, in comparison to an evaluation of r alone, is no more than 5 times the number
of seed vectors. (The factor of 5 is the usual overhead from the reverse mode for a scalar
function.) The space required for storage of the computational graph is no greater than in
the scalar case. As before, we need only store the graph topology information together with
the partial derivative associated with each arc.
The forward- and reverse-mode techniques can be combined to cumulatively reveal
all the elements of J (x). We can choose a set of seed vectors p for the forward mode to
reveal some columns of J , then perform the reverse mode with another set of seed vectors
q to reveal the rows that contain the remaining elements.
Finally, we note that for some algorithms, we do not need full knowledge of the
Jacobian J (x). For instance, iterative methods such as the inexact Newton method for
nonlinear equations (see Section 11.1) require repeated calculation of J (x) p for a succession
of vectors p. Each such matrix–vector product can be computed using the forward mode
by using a single forward sweep, at a similar cost to evaluation of the function alone.
CALCULATING HESSIANS: FORWARD MODE
So far, we have described how the forward and reverse modes can be applied to
obtain first derivatives of scalar and vector functions. We now outline extensions of these
techniques to the computation of the Hessian ∇ 2 f of a scalar function f , and evaluation of
the Hessian–vector product ∇ 2 f (x) p for a given vector p.
Recall that the forward mode makes use of the quantities D p xi , each of which stores
(∇xi )T p for each node i in the computational graph and a given vector p. For a given pair
of seed vectors p and q (both in IRn ) we now define another scalar quantity by
D pq xi p T (∇ 2 xi )q,
(8.36)
for each node i in the computational graph. We can evaluate these quantities during the
forward sweep through the graph, alongside the function values xi and the first-derivative
values D p xi . The initial values of D pq at the independent variable nodes xi , i 1, 2 . . . , n,
will be 0, since the second derivatives of xi are zero at each of these nodes. When the forward
sweep is complete, the value of D pq xi in the rightmost node of the graph will be p T ∇ 2 f (x)q.
The formulae for transformation of the D pq xi variables during the forward sweep
can once again be derived from the chain rule. For instance, if xi is obtained by adding the
values at its two parent nodes, xi x j + xk , the corresponding accumulation operations
on D p xi and D pq xi are as follows:
D p xi D p x j + D p xk ,
D pq xi D pq x j + D pq xk .
(8.37)
213
214
CHAPTER 8.
CALCULATING DERIVATIVES
The other binary operations −, ×, / are handled similarly. If xi is obtained by applying the
unitary transformation L to x j , we have
xi L(x j ),
(8.38a)
D p xi L (x j )(D p x j ),
(8.38b)
D pq xi L (x j )(D p x j )(Dq x j ) + L (x j )D pq x j .
(8.38c)
We see in (8.38c) that computation of D pq xi can rely on the first-derivative quantities
D p xi and Dq xi , so both these quantities must be accumulated during the forward sweep
as well.
We could compute a general dense Hessian by choosing the pairs ( p, q) to be all
possible pairs of unit vectors (e j , ek ), for j 1, 2, . . . , n and k 1, 2, . . . , j, a total of
n(n + 1)/2 vector pairs. (Note that we need only evaluate the lower triangle of ∇ 2 f (x),
because of symmetry.) When we know the sparsity structure of ∇ 2 f (x), we need evaluate
De j ek xi only for the pairs (e j , ek ) for which the ( j, k) component of ∇ 2 f (x) is possibly
nonzero.
The total increase factor for the number of arithmetic operations, compared with the
amount of arithmetic to evaluate f alone, is a small multiple of 1 + n + Nz (∇ 2 f ), where
Nz (∇ 2 f ) is the number of elements of ∇ 2 f that we choose to evaluate. This number reflects
the evaluation of the quantities xi , De j xi ( j 1, 2, . . . , n), and De j ek xi for the Nz (∇ 2 f )
vector pairs (e j , ek ). The “small multiple” results from the fact that the update operations
for D p xi and D pq xi may require a few times more operations than the update operation
for xi alone; see, for example, (8.38). One storage location per node of the graph is required
for each of the 1 + n + Nz (∇ 2 f ) quantities that are accumulated, but recall that storage of
node i can be overwritten once all its children have been evaluated.
When we do not need the complete Hessian, but only a matrix–vector product involving the Hessian (as in the Newton–CG algorithm of Chapter 7), the amount of arithmetic
is, of course, smaller. Given a vector q ∈ IRn , we use the techniques above to compute
the first-derivative quantities De1 xi , . . . Den xi and Dq xi , as well as the second-derivative
quantities De1 q xi , . . . , Den q xi , during the forward sweep. The final node will contain the
quantities
e Tj ∇ 2 f (x) q ∇ 2 f (x)q
j
,
j 1, 2, . . . , n,
which are the components of the vector ∇ 2 f (x)q. Since 2n + 1 quantities in addition to
xi are being accumulated during the forward sweep, the increase factor in the number of
arithmetic operations increases by a small multiple of 2n.
An alternative technique for evaluating sparse Hessians is based on the forwardmode propagation of first and second derivatives of univariate functions. To motivate this
8.2.
AUTOMATIC DIFFERENTIATION
approach, note that the (i, j) element of the Hessian can be expressed as follows:
[∇ 2 f (x)]i j eiT ∇ 2 f (x)e j
1
(ei + e j )T ∇ 2 f (x)(ei + e j ) − eiT ∇ 2 f (x)ei − e Tj ∇ 2 f (x)e j .
2
(8.39)
We can use this interpolation formula to evaluate [∇ 2 f (x)]i j , provided that the second
derivatives D pp xk , for p ei , p e j , p ei + e j , and all nodes xk , have been evaluated
during the forward sweep through the computational graph. In fact, we can evaluate all the
nonzero elements of the Hessian, provided that we use the forward mode to evaluate D p xk
and D pp xk for a selection of vectors p of the form ei + e j , where i and j are both indices
in {1, 2, . . . , n}, possibly with i j.
One advantage of this approach is that it is no longer necessary to propagate “cross
terms” of the form D pq xk for p q (see, for example, (8.37) and (8.38c)). The propagation
formulae therefore simplify somewhat. Each D pp xk is a function of x , D p x , and D pp x
for all parent nodes of node k.
Note, too, that if we define the univariate function ψ by
ψ(t) f (x + t p),
(8.40)
then the values of D p f and D pp f , which emerge at the completion of the forward sweep,
are simply the first two derivatives of ψ evaluated at t 0; that is,
D p f p T ∇ f (x) ψ (t)|t0 ,
D pp f p T ∇ 2 f (x) p ψ (t)|t0 .
Extension of this technique to third, fourth, and higher derivatives is possible. Interpolation formulae analogous to (8.39) can be used in conjunction with higher derivatives
of the univariate functions ψ defined in (8.40), again for a suitably chosen set of vectors p,
where each p is made up of a sum of unit vectors ei . For details, see Bischof, Corliss, and
Griewank [26].
CALCULATING HESSIANS: REVERSE MODE
We can also devise schemes based on the reverse mode for calculating Hessian–
vector products ∇ 2 f (x)q, or the full Hessian ∇ 2 f (x). A scheme for obtaining ∇ 2 f (x)q
proceeds as follows. We start by using the forward mode to evaluate both f and ∇ f (x)T q,
by accumulating the two variables xi and Dq xi during the forward sweep in the manner
described above. We then apply the reverse mode in the normal fashion to the computed
function ∇ f (x)T q. At the end of the reverse sweep, the nodes i 1, 2, . . . , n of the
computational graph that correspond to the independent variables will contain
∂
(∇ f (x)T q) ∇ 2 f (x)q i ,
∂ xi
i 1, 2, . . . , n.
215
216
CHAPTER 8.
CALCULATING DERIVATIVES
The number of arithmetic operations required to obtain ∇ 2 f (x)q by this procedure
increases by only a modest factor, independent of n, over the evaluation of f alone. By
the usual analysis for the forward mode, we see that the computation of f and ∇ f (x)T q
jointly requires a small multiple of the operation count for f alone, while the reverse sweep
introduces a further factor of at most 5. The total increase factor is approximately 12 over the
evaluation of f alone. If the entire Hessian ∇ 2 f (x) is required, we could apply the procedure
just described with q e1 , e2 , . . . , en . This approach would introduce an additional factor
of n into the operation count, leading to an increase of at most 12n over the cost of f alone.
Once again, when the Hessian is sparse with known structure, we may be able to
use graph-coloring techniques to evaluate this entire matrix using many fewer than n seed
vectors. The choices of q are similar to those used for finite-difference evaluation of the
Hessian, described above. The increase in operation count over evaluating f alone is a
multiple of up to 12Nc (∇ 2 f ), where Nc is the number of seed vectors q used in calculating
∇2 f .
CURRENT LIMITATIONS
The current generation of automatic differentiation tools has proved its worth through
successful application to some large and difficult design optimization problems. However,
these tools can run into difficulties with some commonly used programming constructs and
some implementations of computer arithmetic. As an example, if the evaluation of f (x)
depends on the solution of a partial differential equation (PDE), then the computed value
of f may contain truncation error arising from the finite-difference or the finite-element
technique that is used to solve the PDE numerically. That is, we have fˆ(x) f (x) + τ (x),
where fˆ(·) is the computed value of f (·) and τ (·) is the truncation error. Though |τ (x)| is
usually small, its derivative τ (x) may not be, so the error in the computed derivative fˆ (x)
is potentially large. (The finite-difference approximation techniques discussed in Section 8.1
experience the same difficulty.) Similar problems arise when the computer uses piecewise
rational functions to approximate trigonometric functions.
Another source of potential difficulty is the presence of branching in the code to
improve the speed or accuracy of function evaluation in certain domains. A pathological
example is provided by the linear function f (x) x − 1. If we used the following (perverse,
but valid) piece of code to evaluate this function,
if (x 1.0) then f 0.0 else f x − 1.0,
then by applying automatic differentiation to this procedure we would obtain the derivative
value f (1) 0. For a discussion of such issues and an approach to dealing with them, see
Griewank [151, 152].
In conclusion, automatic differentiation should be regarded as a set of increasingly
sophisticated techniques that enhances optimization algorithms, allowing them to be applied
more widely to practical problems involving complicated functions. By providing sensitivity
information, it helps the modeler to extract more information from the results of the
8.2.
AUTOMATIC DIFFERENTIATION
computation. Automatic differentiation should not be regarded as a panacea that absolves
the user altogether from the responsibility of thinking about derivative calculations.
NOTES AND REFERENCES
A comprehensive and authoritative reference on automatic differentiation is the book
of Griewank [152]. The web site www.autodiff.org contains a wealth of current information about theory, software, and applications. A number of edited collections of papers
on automatic differentiation have appeared since 1991; see Griewank and Corliss [153],
Berz et al. [20], and Bücker et al. [40]. An historical paper of note is Corliss and Rall [78],
which includes an extensive bibliography. Software tool development in automatic differentiation makes use not only of forward and reverse modes but also includes “mixed
modes” and “cross-country algorithms” that combine the two approaches; see for example
Naumann [222].
The field of automatic differentiation grew considerably during the 1990s, and and a
number of good software tools appeared. These included ADIFOR [25] and ADIC [28], and
ADOL-C [154]. Tools developed in more recent years include TAPENADE, which accepts
Fortran code through a web server and returns differentiated code; TAF, a commercial tool
that also performs source-to-source automatic differentiation of Fortran codes; OpenAD,
which works with Fortran, C, and C++; and TOMLAB/MAD, which works with MATLAB
code.
The technique for calculating the gradient of a partially separable function was described by Bischof et al. [24], whereas the computation of the Hessian matrix has been
considered by several authors; see, for example, Gay [118].
The work of Coleman and Moré [69] on efficient estimation of Hessians was predated
by Powell and Toint [261], who did not use the language of graph coloring but nevertheless
devised highly effective schemes. Software for estimating sparse Hessians and Jacobians is
described by Coleman, Garbow, and Moré [66, 67]. The recent paper of Gebremedhin,
Manne, and Pothen [120] contains a comprehensive discussion of the application of graph
coloring to both finite difference and automatic differentiation techniques.
✐
EXERCISES
✐
8.1 Show that a suitable value for the perturbation in the central-difference formula
is u1/3 , and that the accuracy achievable by this formula when the values of f contain
roundoff errors of size u is approximately u2/3 . (Use similar assumptions to the ones used
to derive the estimate (8.6) for the forward-difference formula.)
✐
8.2 Derive a central-difference analogue of the Hessian–vector approximation
formula (8.20).
217
218
CHAPTER 8.
CALCULATING DERIVATIVES
✐
8.3 Verify the formula (8.21) for approximating an element of the Hessian using only
function values.
✐
8.4 Verify that if the Hessian of a function f has nonzero diagonal elements, then its
adjacency graph is a subgraph of the intersection graph for ∇ f . In other words, show that
any arc in the adjacency graph also belongs to the intersection graph.
✐
8.5 Draw the adjacency graph for the function f defined by (8.22). Show that the
coloring scheme in which node 1 has one color while nodes 2, 3, . . . , n have another color
is valid. Draw the intersection graph for ∇ f .
✐
8.6 Construct the adjacency graph for the function whose Hessian has the nonzero
structure
⎡
×
⎢
⎢ ×
⎢
⎢ ×
⎢
⎢
⎢ ×
⎢
⎢
⎣
×
×
×
×
×
×
⎤
×
×
×
×
×
×
⎥
⎥
⎥
× ⎥
⎥
⎥,
⎥
⎥
⎥
⎦
×
and find a valid coloring scheme with just four colors.
✐
8.7 Trace the computations performed in the forward mode for the function f (x) in
(8.26), expressing the intermediate derivatives ∇xi , i 4, 5, . . . , 9 in terms of quantities
available at their parent nodes and then in terms of the independent variables x1 , x2 , x3 .
✐
8.8 Formula (8.30) showed the gradient operations associated with scalar division.
Derive similar formulae for the following operations:
(s, t) → s + t
addition;
t → et
exponentiation;
t → tan(t)
tangent;
(s, t) → s .
t
✐
8.9 By calculating the partial derivatives ∂ x j /∂ xi for the function (8.26) from the
expressions (8.27), verify the numerical values for the arcs in Figure 8.3 for the evaluation
point x (1, 2, π/2)T . Work through the remaining details of the reverse sweep process,
indicating the order in which the nodes become finalized.
8.2.
AUTOMATIC DIFFERENTIATION
✐
8.10 Using (8.33) as a guide, describe the reverse sweep operations corresponding to
the following elementary operations in the forward sweep:
xk ← xi x j
xk ← cos(xi )
multiplication;
cosine.
In each case, compare the arithmetic workload in the reverse sweep to the workload required
for the forward sweep.
✐
8.11 Define formulae similar to (8.37) for accumulating the first derivatives D p xi and
the second derivatives D pq xi when xi is obtained from the following three binary operations:
xi x j − xk , xi x j xk , and xi x j /xk .
✐
8.12 By using the definitions (8.28) of D p xi and (8.36) of D pq xi , verify the
differentiation formulae (8.38) for the unitary operation xi L(x j ).
✐ 8.13 Let a ∈ IRn be a fixed vector and define f as f (x) 12 x T x + a T x 2 . Count
the number of operations needed to evaluate f , ∇ f , ∇ 2 f , and the Hessian–vector product
∇ 2 f (x) p for an arbitrary vector p.
219
This is pag
Printer: O
CHAPTER
9
Derivative-Free
Optimization
Many practical applications require the optimization of functions whose derivatives are
not available. Problems of this kind can be solved, in principle, by approximating the
gradient (and possibly the Hessian) using finite differences (see Chapter 8), and using these
approximate gradients within the algorithms described in earlier chapters. Even though
this finite-difference approach is effective in some applications, it cannot be regarded a
general-purpose technique for derivative-free optimization because the number of function
evaluations required can be excessive and the approach can be unreliable in the presence
of noise. (For the purposes of this chapter we define noise to be inaccuracy in the function
evaluation.) Because of these shortcomings, various algorithms have been developed that
9.1.
FINITE DIFFERENCES
AND
NOISE
do not attempt to approximate the gradient. Rather, they use the function values at a set of
sample points to determine a new iterate by some other means.
Derivative-free optimization (DFO) algorithms differ in the way they use the sampled
function values to determine the new iterate. One class of methods constructs a linear
or quadratic model of the objective function and defines the next iterate by seeking to
minimize this model inside a trust region. We pay particular attention to these model-based
approaches because they are related to the unconstrained minimization methods described
in earlier chapters. Other widely used DFO methods include the simplex-reflection method
of Nelder and Mead, pattern-search methods, conjugate-direction methods, and simulated
annealing. In this chapter we briefly discuss these methods, with the exception of simulated
annealing, which is a nondeterministic approach and has little in common with the other
techniques discussed in this book.
Derivative-free optimization methods are not as well developed as gradient-based
methods; current algorithms are effective only for small problems. Although most DFO
methods have been adapted to handle simple types of constraints, such as bounds, the
efficient treatment of general constraints is still the subject of investigation. Consequently,
we limit our discussion to the unconstrained optimization problem
minn f (x).
x∈IR
(9.1)
Problems in which derivatives are not available arise often in practice. The evaluation
of f (x) can, for example, be the result of an experimental measurement or a stochastic
simulation, with the underlying analytic form of f unknown. Even if the objective function
f is known in analytic form, coding its derivatives may be time consuming or impractical.
Automatic differentiation tools (Chapter 8) may not be applicable if f (x) is provided only
in the form of binary computer code. Even when the source code is available, these tools
cannot be applied if the code is written in a combination of languages.
Methods for derivative-free optimization are often used (with mixed success) to
minimize problems with nondifferentiable functions or to try to locate the global minimizer
of a function. Since we do not treat nonsmooth optimization or global optimization in this
book, we will restrict our attention to smooth problems in which f has a continuous
derivative. We do, however, discuss the effects of noise in Sections 9.1 and 9.6.
9.1
FINITE DIFFERENCES AND NOISE
As mentioned above, an obvious DFO approach is to estimate the gradient by using finite
differences and then employ a gradient-based method. This approach is sometimes successful
and should always be considered, but the finite-difference estimates can be inaccurate when
the objective function contains noise. We quantify the effect of noise in this section.
Noise can arise in function evaluations for various reasons. If f (x) depends on a
stochastic simulation, there will be a random error in the evaluated function because of the
221
222
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
finite number of trials in the simulation. When a differential equation solver or some other
complex numerical procedure is needed to calculate f , small but nonzero error tolerances
that are used during the calculations will produce noise in the value of f .
In many applications, then, the objective function f has the form
f (x) h(x) + φ(x),
(9.2)
where h is a smooth function and φ represents the noise. Note that we have written φ to be
a function of x but in practice it need not be. For instance, if the evaluation of f depends
on a simulation, the value of φ will generally differ at each evaluation, even at the same x.
The form (9.2) is, however, useful for illustrating some of the difficulties caused by noise in
gradient estimates and for developing algorithms for derivative-free optimization.
Given a difference interval , recall that the centered finite-difference approximation
(8.7) to the gradient of f at x is defined as follows:
∇ f (x) f (x + ei ) − f (x − ei )
2
,
(9.3)
i1,2,...,n
where ei is the ith unit vector (the vector whose only nonzero element is a 1 in the ith
position). We wish to relate ∇ f (x) to the gradient of the underlying smooth function h(x),
as a function of and the noise level. For this purpose we define the noise level η to be the
largest value of φ in a box of edge length 2 centered at x, that is,
η(x; ) sup
z−x∞ ≤
|φ(z)|.
(9.4)
By applying to the central difference formula (9.3) the argument that led to (8.5), we can
establish the following result.
Lemma 9.1.
Suppose that ∇ 2 h is Lipschitz continuous in a neighborhood of the box {z | z −x∞ ≤ }
with Lipschitz constant L h . Then we have
∇ f (x) − ∇h(x)∞ ≤ L h
2
+
η(x; )
.
(9.5)
Thus the error in the approximation (9.3) comes from both the intrinsic finite difference
approximation error (the O( 2 ) term) and the noise (the η(x; )/ term). If the noise
dominates the difference interval , we cannot expect any accuracy at all in ∇ f (x), so it
will only be pure luck if −∇ f (x) turns out to be a direction of descent for f .
Instead of computing a tight cluster of function values around the current iterate, as
required by a finite-difference approximation to the gradient, it may be preferable to separate
these points more widely and use them to construct a model of the objective function. This
9.2.
MODEL-BASED METHODS
approach, which we consider in the next section and in Section 9.6, may be more robust to
the presence of noise.
9.2
MODEL-BASED METHODS
Some of the most effective algorithms for unconstrained optimization described in the
previous chapters compute steps by minimizing a quadratic model of the objective function
f . The model is formed by using function and derivative information at the current iterate.
When derivatives are not available, we may define the model m k as the quadratic function
that interpolates f at a set of appropriately chosen sample points. Since such a model is
usually nonconvex, the model-based methods discussed in this chapter use a trust-region
approach to compute the step.
Suppose that at the current iterate xk we have a set of sample points Y {y 1 , y 2 , . . . , y q }, with y i ∈ IRn , i 1, 2, . . . , q. We assume that xk is an element of
this set and that no point in Y has a lower function value than xk . We wish to construct a
quadratic model of the form
m k (xk + p) c + g T p + 12 p T Gp.
(9.6)
We cannot define g ∇ f (xk ) and G ∇ 2 f (xk ) because these derivatives are not available.
Instead, we determine the scalar c, the vector g ∈ R n , and the symmetric matrix G ∈ R n×n
by imposing the interpolation conditions
m k (y l ) f (y l ),
l 1, 2, . . . , q.
(9.7)
Since there are 12 (n + 1)(n + 2) coefficients in the model (9.6) (that is, the components of
c, g, and G, taking into account the symmetry of G), the interpolation conditions (9.7)
determine m k uniquely only if
q 12 (n + 1)(n + 2).
(9.8)
In this case, (9.7) can be written as a square linear system of equations in the coefficients of
the model. If we choose the interpolation points y 1 , y 2 , . . . , y q so that this linear system is
nonsingular, the model m k will be uniquely determined.
Once m k has been formed, we compute a step p by approximately solving the trustregion subproblem
min m k (xk + p),
p
subject to p2 ≤ ,
(9.9)
for some trust-region radius > 0. We can use one of the techniques described in Chapter 4
to solve this subproblem. If xk + p gives a sufficient reduction in the objective function,
223
224
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
the new iterate is defined as xk+1 xk + p, the trust region radius is updated, and a
new iteration commences. Otherwise the step is rejected, and the interpolation set Y may
be improved or the trust region shrunk.
To reduce the cost of the algorithm, we update the model m k at every iteration,
rather than recomputing it from scratch. In practice, we choose a convenient basis for the
space of quadratic polynomials, the most common choices being Lagrange and Newton
polynomials. The properties of these bases can be used both to measure appropriateness of
the sample set Y and to change this set if necessary. A complete algorithm that treats all
these issues effectively is far more complicated than the quasi-Newton methods discussed
in Chapter 6. Consequently, we will provide only a broad outline of model-based DFO
methods.
As is common in trust-region algorithms, the step-acceptance and trust-region update
strategies are based on the ratio between the actual reduction in the function and the
reduction predicted by the model, that is,
ρ
f (xk ) − f (xk+ )
,
m k (xk ) − m k (xk+ )
(9.10)
where xk+ denotes the trial point. Throughout this section, the integer q is defined
by (9.8).
Algorithm 9.1 (Model-Based Derivative-Free Method).
Choose an interpolation set Y {y 1 , y 2 , . . . , y q } such that the linear system defined
by (9.7) is nonsingular, and select x0 as a point in this set such that f (x0 ) ≤ f (y i ) for all
y i ∈ Y . Choose an initial trust region radius 0 , a constant η ∈ (0, 1), and set k ← 0.
repeat until a convergence test is satisfied:
Form the quadratic model m k (xk + p) that satisfies the interpolation
conditions (9.7);
Compute a step p by approximately solving subproblem (9.9);
Define the trial point as xk+ xk + p;
Compute the ratio ρ defined by (9.10);
if ρ ≥ η
Replace an element of Y by xk+ ;
Choose k+1 ≥ k ;
Set xk+1 ← xk+ ;
Set k ← k + 1 and go to the next iteration;
else if the set Y need not be improved
Choose k+1 < k ;
Set xk+1 ← xk ;
Set k ← k + 1 and go to the next iteration;
end (if)
9.2.
MODEL-BASED METHODS
Invoke a geometry-improving procedure to update Y :
at least one of the points in Y is replaced by some other point,
with the goal of improving the conditioning of (9.7);
Set k+1 ← k ;
Choose x̂ as an element in Y with lowest function value;
Set xk+ ← x̂ and recompute ρ by (9.10);
if ρ ≥ η
Set xk+1 ← xk+ ;
else
Set xk+1 ← xk ;
end (if)
Set k ← k + 1;
end (repeat)
The case of ρ ≥ η, in which we obtain sufficient reduction in the merit function, is
the simplest. In this case we always accept the trial point xk+ as the new iterate, include xk+
in Y , and remove an element from Y .
When sufficient reduction is not achieved (ρ < η), we look at two possible causes:
inadequacy of the interpolation set Y and a trust region that is too large. The first cause
can arise when the iterates become restricted to a low-dimensional surface of IRn that does
not contain the solution. The algorithm could then be converging to a minimizer in this
subset. Behavior such as this can be detected by monitoring the conditioning of the linear
system defined by the interpolation conditions (9.7). If the condition number is too high,
we change Y to improve it, typically by replacing one element of Y with a new element
so as to move the interpolation system (9.7) as far away from singularity as possible. If Y
seems adequate, we simply decrease the trust region radius , as is done in the methods of
Chapter 4.
A good initial choice for Y is given by the vertices and the midpoints of the edges of a
simplex in IRn .
The use of quadratic models limits the size of problems that can be solved in
practice. Performing O(n 2 ) function evaluations just to start the algorithm is onerous,
even for moderate values of n (say, n 50). In addition, the cost of the iteration is
high. Even by updating the model m k at every iteration, rather than recomputing it
from scratch, the number of operations required to construct m k and compute a step
is O(n 4 ) [257].
To alleviate these drawbacks, we can replace the quadratic model by a linear model in
which the matrix G in (9.6) is set to zero. Since such a model contains only n + 1 parameters,
we need to retain only n + 1 interpolation points in the set Y , and the cost of each iteration is
O(n 3 ). Algorithm 9.1 can be applied with little modification when the model is linear, but it
is not rapidly convergent because linear models cannot represent curvature of the problem.
Therefore, some model-based algorithms start with n + 1 initial points and compute steps
225
226
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
using a linear model, but after q 12 (n + 1)(n + 2) function values become available, they
switch to using quadratic models.
INTERPOLATION AND POLYNOMIAL BASES
We now consider in more detail how to form a model of the objective function using
interpolation techniques. We begin by considering a linear model of the form
m k (xk + p) f (xk ) + g T p.
(9.11)
To determine the vector g ∈ IRn , we impose the interpolation conditions m k (y l ) f (y l ),
l 1, 2, . . . , n, which can be written as
(s l )T g f (y l ) − f (xk ),
l 1, 2, . . . , n,
(9.12)
where
s l yl − xk ,
l 1, 2, . . . , n.
(9.13)
Conditions (9.12) represent a linear system of equations in which the rows of the coefficient
matrix are given by the vectors (s l )T . It follows that the model (9.11) is determined uniquely
by (9.12) if and only if the interpolation points {y 1 , y 2 , . . . , y n } are such that the set
{s l : l 1, 2, . . . , n} is linearly independent. If this condition holds, the simplex formed by
the points xk , y 1 , y 2 , . . . , y n is said to be nondegenerate.
Let us now consider how to construct a quadratic model of the form (9.6), with
f f (xk ). We rewrite the model as
m k (xk + p) f (xk ) + g T p +
G i j pi p j +
i< j
1
2
G ii pi2
(9.14)
i
def
f (xk ) + ĝ T p̂,
(9.15)
where we have collected the elements of g and G in the (q − 1)-vector of unknowns
-T
,
,
ĝ ≡ g T , {G i j }i< j , √12 G ii
(9.16)
and where the (q − 1)-vector p̂ is given by
-T
,
p̂ ≡ p T , { pi p j }i< j , √12 pi2
.
The model (9.15) has the same form as (9.11), and the determination of the vector of
unknown coefficients ĝ can be done as in the linear case.
9.2.
MODEL-BASED METHODS
Multivariate quadratic functions can be represented in various ways. The monomial
basis (9.14) has the advantage that known structure in the Hessian can be imposed easily by
setting appropriate elements in G to zero. Other bases are, however, more convenient when
one is developing mechanisms for avoiding singularity of the system (9.7).
q
We denote by {φi (·)}i1 a basis for the linear space of n-dimensional quadratic
functions. The function (9.6) can therefore be expressed as
m k (x) q
αi φi (x),
i1
for some coefficients αi . The interpolation set Y {y 1 , y 2 , . . . , y q } determines the
coefficients αi uniquely if the determinant
⎛
φ1 (y 1 )
⎜
def
..
δ(Y ) det ⎜
.
⎝
···
φq (y 1 )
···
⎞
φ1 (y q )
⎟
..
⎟
.
⎠
(9.17)
φq (y q )
is nonzero.
As model-based algorithms iterate, the determinant δ(Y ) may approach zero, leading
to numerical difficulties or even failure. Several algorithms therefore contain a mechanism
for keeping the interpolation points well placed. We now describe one of those mechanisms.
UPDATING THE INTERPOLATION SET
Rather than waiting until the determinant δ(Y ) becomes smaller than a threshold,
we may invoke a geometry-improving procedure whenever a trial point does not provide
sufficient decrease in f . The goal in this case is to replace one of the interpolation points
so that the determinant (9.17) increases in magnitude. To guide us in this exchange, we use
the following property of δ(Y ), which we state in terms of Lagrange functions.
For every y ∈ Y , we define the Lagrangian function L(·, y) to be a polynomial of
degree at most 2 such that L(y, y) 1 and L( ŷ, y) 0 for ŷ y, ŷ ∈ Y . Suppose that the
set Y is updated by removing a point y− and replacing it by some other point y+ , to give the
new set Y + . One can show that (after a suitable normalization and given certain conditions
[256])
|δ(Y + )| ≤ |L(y+ , y− )| |δ(Y )|.
(9.18)
Algorithm 9.1 can make good use of this inequality to update the interpolation set.
Consider first the case in which trial point x + provides sufficient reduction in the
objective function (ρ ≥ η). We include x + in Y and remove another point y− from Y .
227
228
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
Motivated by (9.18), we select the outgoing point as follows:
y− arg max |L(x + , y)|.
y∈Y
Next, let us consider the case in which the reduction in f is not sufficient (ρ < η).
We first determine whether the set Y should be improved, and for this purpose we use
the following rule. We consider Y to be adequate at the current iterate xk if for all y i ∈ Y
such that xk − y i ≤ we have that |δ(Y )| cannot be doubled by replacing one of these
interpolation points y i with any point y inside the trust region. If Y is adequate but the
reduction in f was not sufficient, we decrease the trust-region radius and begin a new
iteration.
If Y is inadequate, the geometry-improving mechanism is invoked. We choose a point
y− ∈ Y and replace it by some other point y + that is chosen solely with the objective
of improving the determinant (9.17). For every point y i ∈ Y , we define its potential
replacement yri as
yri arg max |L(y, y i )|.
y−xk ≤
The outgoing point y− is selected as the point for which |L(yri , y i )| is maximized over all
indices y i ∈ Y .
Implementing these rules efficiently in practice is not simple, and one must also
consider several possible difficulties we have not discussed; see [76]. Strategies for improving
the position of the interpolation set are the subject of ongoing investigation and new
developments are likely in the coming years.
A METHOD BASED ON MINIMUM-CHANGE UPDATING
We now consider a method that be viewed as an extension of the quasi-Newton
approach discussed in Chapter 6. The method uses quadratic models but requires only
O(n 3 ) operations per iteration, substantially fewer than the O(n 4 ) operations required by
the methods described above. To achieve this economy, the method retains only O(n) points
for the interpolation conditions (9.7) and absorbs the remaining degrees of freedom in the
model (9.6) by requiring that the Hessian of the model change as little as possible from
one iteration to the next. This least-change property is one of the key ingredients in quasiNewton methods, the other ingredient being the requirement that the model interpolate
the gradient ∇ f at the two most recent points. The method we describe now combines the
least-change property with interpolation of function values.
At the kth iteration of the algorithm, a new quadratic model m k+1 of the form (9.6)
is constructed after taking a step from xk to xk+1 . The coefficients f k+1 , gk+1 , G k+1 of the
9.3.
COORDINATE
AND
PATTERN-SEARCH METHODS
model m k+1 are determined as the solution of the problem
min
f,g,G
subject to
G − G k 2F
(9.19a)
G symmetric
m(y l ) f (y l )
l 1, 2, . . . , q̂,
(9.19b)
where · F denotes the Frobenius norm (see (A.9)), G k is the Hessian of the previous
model m k , and q̂ is an integer comparable to n. One can show that the integer q̂ must be
chosen larger than n + 1 to guarantee that G k+1 is not equal to G k . An appropriate value in
practice is q̂ 2n + 1; for this choice the number of interpolation points is roughly twice
that used for linear models.
Problem (9.19) is an equality-constrained quadratic program whose KKT conditions
can be expressed as a system of equations. Once the model m k+1 is determined, we compute
a new step by solving a trust-region problem of the form (9.9). In this approach, too,
it is necessary to ensure that the geometry of the interpolation set Y is adequate. We
therefore impose two minimum requirements. First, the set Y should be such that the
equations (9.19b) can be satisfied for any right-hand side. Second, the points y i should
not all lie in a hyperplane. If these two conditions hold, problem (9.19) has a unique
solution.
A practical algorithm based on the subproblem (9.19) resembles Algorithm 9.1 in that
it contains procedures both for generating new iterates and for improving the geometry
of the set Y . The implementation described in [260] contains other features to ensure that
the interpolation points are well separated and that steps are not too small. A strength of
this method is that it requires only O(n) interpolation points to start producing productive
steps. In practice the method often approaches a solution with fewer than 12 (n + 1)(n + 2)
function evaluations. However, since this approach has been developed only recently, there
is insufficient numerical experience to assess its full potential.
9.3
COORDINATE AND PATTERN-SEARCH METHODS
Rather than constructing a model of f explicitly based on function values, coordinate search
and pattern-search methods look along certain specified directions from the current iterate
for a point with a lower function value. If such a point is found, they step to it and repeat the
process, possibly modifying the directions of search for the next iteration. If no satisfactory
new point is found, the step length along the current search directions may be adjusted, or
new search directions may be generated.
We describe first a simple approach of this type that has been used often in practice.
We then consider a generalized approach that is potentially more efficient and has stronger
theoretical properties.
229
230
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
x*
x0
x
1
Figure 9.1
Coordinate search method makes slow
progress on this function of two variables.
COORDINATE SEARCH METHOD
The coordinate search method (also known as the coordinate descent method or
the alternating variables method) cycles through the n coordinate directions e1 , e2 , . . . , en ,
obtaining new iterates by performing a line search along each direction in turn. Specifically,
at the first iteration, we fix all components of x except the first one x1 and find a new value
of this component that minimizes (or at least reduces) the objective function. On the next
iteration, we repeat the process with the second component x2 , and so on. After n iterations,
we return to the first variable and repeat the cycle. Though simple and somewhat intuitive,
this method can be quite inefficient in practice, as we illustrate in Figure 9.1 for a quadratic
function in two variables. Note that after a few iterations, neither the vertical (x2 ) nor the
horizontal (x1 ) move makes much progress toward the solution at each iteration.
In general, the coordinate search method can iterate infinitely without ever approaching a point where the gradient of the objective function vanishes, even when exact line
searches are used. (By contrast, as we showed in Section 3.2, the steepest descent method
produces a sequence of iterates {xk } for which ∇ f k → 0, under reasonable assumptions.)
In fact, a cyclic search along any set of linearly independent directions does not guarantee
global convergence [243]. Technically speaking, this difficulty arises because the steepest descent search direction −∇ f k may become more and more perpendicular to the coordinate
search direction. In such circumstances, the Zoutendijk condition (3.14) is satisfied because
cos θk approaches zero rapidly, even when ∇ f k does not approach zero.
When the coordinate search method does converge to a solution, it often converges
much more slowly than the steepest descent method, and the difference between the two
approaches tends to increase with the number of variables. However, coordinate search may
9.3.
COORDINATE
AND
PATTERN-SEARCH METHODS
still be useful because it does not require calculation of the gradient ∇ f k , and the speed
of convergence can be quite acceptable if the variables are loosely coupled in the objective
function f .
Many variants of the coordinate search method have been proposed, some of which
allow a global convergence property to be proved. One simple variant is a “back-and-forth”
approach in which we search along the sequence of directions
e1 , e2 , . . . , en−1 , en , en−1 , . . . , e2 , e1 , e2 , . . .
(repeats).
Another approach, suggested by Figure 9.1, is first to perform a sequence of coordinate
descent steps and then search along the line joining the first and last points in the cycle. Several
algorithms, such as that of Hooke and Jeeves, are based on these ideas; see Fletcher [101]
and Gill, Murray, and Wright [130].
The pattern-search approach, described next, generalizes coordinate search in that it
allows the use of a richer set of search directions at each iteration.
PATTERN-SEARCH METHODS
We consider pattern-search methods that choose a certain set of search directions
at each iterate and evaluate f at a given step length along each of these directions. These
candidate points form a “frame,” or “stencil,” around the current iterate. If a point with a
significantly lower function value is found, it is adopted as the new iterate, and the center
of the frame is shifted to this new point. Whether shifted or not, the frame may then be
altered in some way (the set of search directions may be changed, or the step length may
grow or shrink), and the process repeats. For certain methods of this type it is possible
to prove global convergence results—typically, that there exists a stationary accumulation
point.
The presence of noise or other forms of inexactness in the function values may affect
the performance of pattern-search algorithms and certainly impacts the convergence theory.
Nonsmoothness may also cause undesirable behavior, as can be shown by simple examples,
although satisfactory convergence is often observed on nonsmooth problems.
To define pattern-search methods, we introduce some notation. For the current iterate
xk , we define Dk to be the set of possible search directions and γk to be the line search
parameter. The frame consists of the points xk + γk pk , for all pk ∈ Dk . When one of the
points in the frame yields a significant decrease in f , we take the step and may also increase
γk , so as to expand the frame for the next iteration. If none of the points in the frame has a
significantly better function value than f k , we reduce γk (contract the frame), set xk+1 xk ,
and repeat. In either case, we may change the direction set Dk prior to the next iteration,
subject to certain restrictions.
A more precise description of the algorithm follows.
231
232
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
Algorithm 9.2 (Pattern-Search).
Given convergence tolerance γtol , contraction parameter θmax ,
sufficient decrease function ρ : [0, ∞) → IR with ρ(t) an increasing
function of t and ρ(t)/t → 0 as t ↓ 0;
Choose initial point x0 , initial step length γ0 > γtol , initial direction set D0 ;
for k 1, 2, . . .
if γk ≤ γtol
stop;
if f (xk + γk pk ) < f (xk ) − ρ(γk ) for some pk ∈ Dk
Set xk+1 ← xk + γk pk for some such pk ;
Set γk+1 ← φk γk for some φk ≥ 1; (∗ increase step length ∗ )
else
Set xk+1 ← xk ;
Set γk+1 ← θk γk , where 0 < θk ≤ θmax < 1;
end (if)
end (for)
A wise choice of the direction set Dk is crucial to the practical behavior of this approach
and to the theoretical results that can be proved about it. A key condition is that at least one
direction in this set should give a direction of descent for f whenever ∇ f (xk ) 0 (that is,
whenever xk is not a stationary point). To make this condition specific, we refer to formula
(3.12), where we defined the angle between a possible search direction d and the gradient
∇ f k as follows:
cos θ −∇ f kT p
.
∇ f k p
(9.20)
Recall from Theorem 3.2 that global convergence of a line-search method to a stationary
point of f could be ensured if the search direction d at each iterate xk satisfied cos θ ≥ δ,
for some constant δ > 0, and if the line search parameter satisfied certain conditions. In
the same spirit, we choose Dk so that at least one direction p ∈ Dk will yield cos θ > δ,
regardless of the value of ∇ f k . This condition is as follows:
def
κ(Dk ) minn max
v∈IR p∈Dk
vT p
≥ δ.
v p
(9.21)
A second condition on Dk is that the lengths of the vectors in this set are all roughly
similar, so that the diameter of the frame formed by this set is captured adequately by the
step length parameter γk . Thus, we impose the condition
βmin ≤ p ≤ βmax , for all p ∈ Dk ,
(9.22)
for some positive constants βmin and βmax and all k. If the conditions (9.21) and (9.22) hold,
9.3.
COORDINATE
AND
PATTERN-SEARCH METHODS
we have for any k that
−∇ f kT p ≥ κ(Dk )∇ f k p ≥ δβmin ∇ f k , for some p ∈ Dk .
Examples of sets Dk that satisfy the properties (9.21) and (9.22) include the coordinate
direction set
{e1 , e2 , . . . , en , −e1 , −e2 , . . . , −en },
(9.23)
and the set of n + 1 vectors defined by
pi 1
e − ei , i 1, 2, . . . , n;
2n
pn+1 1
e,
2n
(9.24)
where e (1, 1, . . . , 1)T . For n 3 these direction sets are sketched in Figure 9.2.
The coordinate descent method described above is similar to the special case of
Algorithm 9.2 obtained by setting Dk {ei , −ei } for some i 1, 2, . . . , n at each iteration.
Note that for this choice of Dk , we have κ(Dk ) 0 for all k. Hence, as noted above, cos θ
can be arbitrarily close to zero at each iteration.
Often, the directions that satisfy the properties (9.21) and (9.22) form only a subset of
the direction set Dk , which may contain other directions as well. These additional directions
could be chosen heuristically, according to some knowledge of the function f and its
scaling, or according to experience on previous iterations. They could also be chosen as
linear combinations of the core set of directions (the ones that ensure δ > 0).
Note that Algorithm 9.2 does not require us to choose the point xk + γk pk , pk ∈ Dk ,
with the smallest objective value. Indeed, we may save on function evaluations by not
evaluating f at all points in the frame, but rather performing the evaluations one at a time
and accepting the first candidate point that satisfies the sufficient decrease condition.
e3
p4
e2
p
1
1
e1
2
p
2
p
3
3
Figure 9.2 Generating search sets in IR3 : coordinate direction set (left) and simplex
set (right).
233
234
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
Another important detail in the implementation of Algorithm 9.2 is the choice of
sufficient decrease function ρ(t). If ρ(·) is chosen to be identically zero, then any candidate
point that produces a decrease in f is acceptable as a new iterate. As we have seen in Chapter 3,
such a weak condition does not lead to strong global convergence results in general. A more
appropriate choice might be ρ(t) Mt 3/2 , where M is some positive constant.
9.4
A CONJUGATE-DIRECTION METHOD
We have seen in Chapter 5 that the minimizer of a strictly convex quadratic function
f (x) 12 x T Ax − b T x
(9.25)
can be located by performing one-dimensional minimizations along a set of n conjugate
directions. These directions were defined in Chapter 5 as a linear combination of gradients.
In this section, we show how to construct conjugate directions using only function values,
and we therefore devise an algorithm for minimizing (9.25) that requires only function
value calculations. Naturally, we also consider an extension of this approach to the case of a
nonlinear objective f .
We use the parallel subspace property, which we describe first for the case n 2.
Consider two parallel lines l1 (α) x1 + αp and l2 (α) x2 + αp, where x1 , x2 , and p are
given vectors in IR2 and α is the scalar parameter that defines the lines. We show below that
if x1∗ and x2∗ denote the minimizers of f (x) along l1 and l2 , respectively, then x1∗ − x2∗ is
conjugate to p. Hence, if we perform a one-dimensional minimization along the line joining
x1∗ and x2∗ , we will reach the minimizer of f , because we have successively minimized along
the two conjugate directions p and x2∗ − x1∗ . This process is illustrated in Figure 9.3.
This observation suggests the following algorithm for minimizing a two-dimensional
quadratic function f . We choose a set of linearly independent directions, say the coordinate
directions e1 and e2 . From any initial point x0 , we first minimize f along e2 to obtain the
point x1 . We then perform successive minimizations along e1 and e2 , starting from x1 , to
obtain the point z. It follows from the parallel subspace property that z − x1 is conjugate
to e2 because x1 and z are minimizers along two lines parallel to e2 . Thus, if we perform a
one-dimensional search from x1 along the direction z −x1 , we will locate the minimizer of f .
We now state the parallel subspace minimization property in its most general form.
Suppose that x1 , x2 are two distinct points in IRn and that { p1 , p2 , . . . , pl } is a set of linearly
independent directions in IRn . Let us define the two parallel linear varieties
S1 x1 +
S2 x2 +
l
i1
l
i1
αi pi | αi ∈ IR, i 1, 2, . . . , l ,
αi pi | αi ∈ IR, i 1, 2, . . . , l .
9.4.
A CONJUGATE-DIRECTION METHOD
x *2
l2
x*
Figure 9.3
Geometric construction of
conjugate directions. (The
minimizer of f is denoted by
x ∗ .)
x1*
l1
If we denote the minimizers of f on S1 and S2 by x1∗ and x2∗ , respectively, then x2∗ − x1∗ is
conjugate to p1 , p2 , . . . , pl . It is easy to verify this claim. By the minimization property, we
have that
∂ f (x1∗ + αi pi ) ∇ f (x1∗ )T pi 0,
i 1, 2, . . . , l,
∂αi
αi 0
and similarly for x2 . Therefore we have from (9.25) that
0 (∇ f (x1∗ ) − ∇ f (x2∗ ))T pi
(Ax1∗ − b − Ax2∗ + b)T pi
(x1∗ − x2∗ )T Api ,
i 1, 2, . . . , l.
(9.26)
We now consider the case n 3 and show how the parallel subspace property
can be used to generate a set of three conjugate directions. We choose a set of linearly
independent directions, say e1 , e2 , e3 . From any starting point x0 we first minimize f along
the last direction e3 to obtain a point x1 . We then perform three successive one-dimensional
minimizations, starting from x1 , along the directions e1 , e2 , e3 and denote the resulting
point by z. Next, we minimize f along the direction p1 z − x1 to obtain x2 . As noted
earlier, p1 z − x1 is conjugate to e3 . We note also that x2 is the minimizer of f on the set
S1 {y + α1 e3 + α2 p1 | α1 ∈ IR, α2 ∈ IR}, where y is the intermediate point obtained after
minimizing along e1 and e2 .
A new iteration now commences. We discard e1 and define the new set of search
directions as e2 , e3 , p1 . We perform one-dimensional minimizations along e2 , e3 , p1 , starting
235
236
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
from x2 , to obtain the point ẑ. Note that ẑ can be viewed as the minimizer of f on the set
S2 { ŷ + α1 e3 + α2 p1 | α1 ∈ IR, α2 ∈ IR}, for some intermediate point ŷ. Therefore, by
applying the parallel subspace minimization property to the sets S1 and S2 just defined, we
have that p2 ẑ − x2 is conjugate to both e3 and p1 . We then minimize f along p2 to
obtain a point x3 , which is the minimizer of f . This procedure thus generates the conjugate
directions e3 , p1 , p2 .
We can now state the general algorithm, which consists of an inner and an outer
iteration. In the inner iteration, n one-dimensional minimizations are performed along a set
of linearly independent directions. Upon completion of the inner iteration, a new conjugate
direction is generated, which replaces one of the previously stored search directions.
Algorithm 9.3 (DFO Method of Conjugate Directions).
Choose an initial point x0 and set pi ei , for i 1, 2, . . . , n;
Compute x1 as the minimizer of f along the line x0 + αpn ;
Set k ← 1.
repeat until a convergence test is satisfied
Set z 1 ← xk ;
for j 1, 2, . . . , n
Calculate α j so that f (z j + α j p j ) is minimized;
Set z j+1 ← z j + α j p j ;
end (for)
Set p j ← p j+1 for j 1, 2, . . . , n − 1 and pn ← z n+1 − z 1 ;
Calculate αn so that f (z n+1 + αn pn ) is minimized;
Set xk+1 ← z n+1 + αn pn ;
Set k ← k + 1;
end (repeat)
The line searches can be performed by quadratic interpolation using three function
values along each search direction. Since the restriction of (9.25) to a line is a (strictly
convex) quadratic, the interpolating quadratic matches it exactly, and the one-dimensional
minimizer can easily be computed. Note that at the end of (the outer) iteration k, the
directions pn−k , pn−k+1 , . . . , pn are conjugate by the property mentioned above. Thus the
algorithm terminates at the minimizer of (9.25) after n − 1 iterations, provided none of the
conjugate directions is zero. Unfortunately, this possibility cannot be ruled out, and some
safeguards described below must be incorporated to improve robustness. In the (usual)
case that Algorithm 9.3 terminates after n − 1 iterations, it will perform O(n 2 ) function
evaluations.
Algorithm 9.3 can be extended to minimize nonquadratic objective functions. The
only change is in the line search, which must be performed approximately, using interpolation. Because of the possible nonconvexity, this one-dimensional search must be done with
care; see Brent [39] for a treatment of this subject. Numerical experience indicates that this
9.4.
A CONJUGATE-DIRECTION METHOD
extension of Algorithm 9.3 performs adequately for small-dimensional problems but that
sometimes the directions { pi } tend to become linearly dependent. Several modifications of
the algorithm have been proposed to guard against this possibility. One such modification
measures the degree to which the directions { pi } are conjugate. To do so, we define the
scaled directions
p̂i 0
pi
piT Api
,
i 1, 2, . . . , n.
(9.27)
One can show [239] that the quantity
| det( p̂1 , p̂2 , . . . , p̂n )|
(9.28)
is maximized if and only if the vectors pi are conjugate with respect to A. This result suggests
that we should not replace one of the existing search directions in the set { p1 , p2 , . . . , pn }
by the most recently generated conjugate direction if this action causes the quantity (9.28)
to decrease.
Procedure 9.4 implements this strategy for the case of the quadratic objective function
(9.25). Some algebraic manipulations (which we do not present here) show that we can
compute the scaled directions p̂i without using the Hessian A because the terms piT Api
are available from the line search along pi . Further, only comparisons using computed
function values are needed to ensure that (9.28) does not increase. The following procedure is invoked immediately after the execution of the inner iteration (or for-loop) of
Algorithm 9.3.
Procedure 9.4 (Updating of the Set of Directions).
Find the integer m ∈ {1, 2, . . . , n} such that ψm f (xm−1 ) − f (xm )
is maximized;
Let f 1 f (z 1 ), f 2 f (z n+1 ), and f 3 f (2z n+1 − z 1 );
if f 3 ≥ f 1 or ( f 1 − 2 f 2 + f 3 )( f 1 − f 2 − ψm )2 ≥ 12 ψm ( f 1 − f 3 )2
Keep the set p1 , p2 , . . . , pn unchanged and set xk+1 ← z n+1 ;
else
Set p̂ ← z n+1 − z 1 and calculate α̂ so that f (z n+1 + α̂ p̂) is minimized;
Set xk+1 ← z n+1 + α p̂;
Remove pm from the set of directions and add p̂ to this set;
end (if)
This procedure can be applied to general objective functions by implementing inexact
one-dimensional line searches. The resulting conjugate-gradient method has been found to
be useful for solving small dimensional problems.
237
238
CHAPTER 9.
9.5
DERIVATIVE-FREE OPTIMIZATION
NELDER–MEAD METHOD
The Nelder–Mead simplex-reflection method has been a popular DFO method since its
introduction in 1965 [223]. It takes its name from the fact that at any stage of the algorithm,
we keep track of n + 1 points of interest in IRn , whose convex hull forms a simplex. (The
method has nothing to do with the simplex method for linear programming discussed
in Chapter 13.) Given a simplex S with vertices {z 1 , z 2 , . . . , z n+1 }, we can define an associated matrix V (S) by taking the n edges along V from one of its vertices (z 1 , say), as
follows:
V (S) [z 2 − z 1 , z 3 − z 1 , . . . , z n+1 − z 1 ] .
The simplex is said to be nondegenerate or nonsingular if V is a nonsingular matrix.
(For example, a simplex in IR3 is nondegenerate if its four vertices are not
coplanar.)
In a single iteration of the Nelder–Mead algorithm, we seek to remove the vertex with
the worst function value and replace it with another point with a better value. The new
point is obtained by reflecting, expanding, or contracting the simplex along the line joining
the worst vertex with the centroid of the remaining vertices. If we cannot find a better point
in this manner, we retain only the vertex with the best function value, and we shrink the
simplex by moving all other vertices toward this value.
We specify a single step of the algorithm after some defining some notation. The
n + 1 vertices of the current simplex are denoted by {x1 , x2 , . . . , xn+1 }, where we choose the
ordering so that
f (x1 ) ≤ f (x2 ) ≤ · · · ≤ f (xn+1 ).
The centroid of the best n points is denoted by
x̄ n
xi .
i1
Points along the line joining x̄ and the “worst” vertex xn+1 are denoted by
x̄(t) x̄ + t(xn+1 − x̄).
Procedure 9.5 (One Step of Nelder–Mead Simplex).
Compute the reflection point x̄(−1) and evaluate f −1 f (x̄(−1));
if f (x1 ) ≤ f −1 < f (xn )
(∗ reflected point is neither best nor worst in the new simplex ∗ )
replace xn+1 by x̄(−1) and go to next iteration;
else if f −1 < f (x1 )
9.5.
NELDER–MEAD METHOD
(∗ reflected point is better than the current best; try to
go farther along this direction ∗ )
Compute the expansion point x̄(−2) and evaluate f −2 f (x̄(−2));
if f −2 < f −1
replace xn+1 by x−2 and go to next iteration;
else
replace xn+1 by x−1 and go to next iteration;
else if f −1 ≥ f (xn )
(∗ reflected point is still worse than xn ; contract ∗ )
if f (xn ) ≤ f −1 < f (xn+1 )
(∗ try to perform “outside” contraction ∗ )
evaluate f −1/2 x̄(−1/2);
if f −1/2 ≤ f −1
replace xn+1 by x−1/2 and go to next iteration;
else
(∗ try to perform “inside” contraction ∗ )
evaluate f 1/2 x̄(1/2);
if f 1/2 < f n+1
replace xn+1 by x1/2 and go to next iteration;
∗
( neither outside nor inside contraction was acceptable;
shrink the simplex toward x1 ∗ )
replace xi ← (1/2)(x1 + xi ) for i 2, 3, . . . , n + 1;
Procedure 9.5 is illustrated on a three-dimensional example in Figure 9.4. The worst
current vertex is x3 , and the possible replacement points are x̄(−1), x̄(−2), x̄(− 12 ), x̄( 12 ). If
none of the replacement points proves to be satisfactory, the simplex is shrunk to the smaller
triangle indicated by the dotted line, which retains the best vertex x1 . The scalars t used
in defining the candidate points x̄(t) have been assigned the specific (and standard) values
−1, −2, − 12 , and 12 in our description above. Different choices are also possible, subject to
certain restrictions.
Practical performance of the Nelder–Mead algorithm is often reasonable, though
stagnation has been observed to occur at nonoptimal points. Restarting can be used when
stagnation is detected; see Kelley [178]. Note that unless the final shrinkage step is performed,
the average function value
1 f (xi )
n + 1 i1
n+1
(9.29)
will decrease at each step. When f is convex, even the shrinkage step is guaranteed not to
increase the average function value.
239
240
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
x2
x (−2)
x (−1)
x (−1/2)
x (1/2)
x3
x1
Figure 9.4 One step of the Nelder–Mead simplex method in IR3 , showing current
simplex (solid triangle with vertices x1 , x2 , x3 ), reflection point x̄(−1), expansion
point x̄(−2), inside contraction point x̄( 12 ), outside contraction point x̄(− 12 ), and
shrunken simplex (dotted triangle).
A limited amount of convergence theory has been developed for the Nelder–Mead
method in recent years; see, for example, Kelley [179] and Lagarias et al. [186].
9.6 IMPLICIT FILTERING
We now describe an algorithm designed for functions whose evaluations are modeled by
(9.2), where h is smooth. This implicit filtering approach is, in its simplest form, a variant of
the steepest descent algorithm with line search discussed in Chapter 3, in which the gradient
∇ f k is replaced by a finite difference estimate such as (9.3), with a difference parameter
that may not be particularly small.
Implicit filtering works best on functions for which the noise level decreases as the
iterates approach a solution. This situation may occur when we have control over the noise
level, as is the case when f is obtained by solving a differential equation to a user-specified
tolerance, or by running a stochastic simulation for a user-specified number of trials (where
an increase in the number of trials usually produces a decrease in the noise). The implicit
filtering algorithm decreases systematically (but, one hopes, not as rapidly as the decay in
error) so as to maintain reasonable accuracy in ∇ f (x), given the noise level at the current
value of x. For each value of , it performs an inner loop that is simply an Armijo line search
using the search direction −∇ f (x). If the inner loop is unable to find a satisfactory step
length after backtracking at least amax times, we return to the outer loop, choose a smaller
value of , and repeat. A formal specification follows.
9.6.
IMPLICIT FILTERING
Algorithm 9.6 (Implicit Filtering).
Choose a sequence { k } ↓ 0, Armijo parameters c and ρ in (0, 1),
maximum backtracking parameter amax ;
Set k ← 1, Choose initial point x x0 ;
repeat
increment k ← false;
repeat
Compute
f (x)
and ∇ k f (x);
if ∇ k f (x) ≤ k
increment k ← true;
else
Find the smallest integer m between 0 and amax such that
2
f x − ρ m ∇ k f (x) ≤ f (x) − cρ m ∇ k f (x)2 ;
if no such m exists
increment k ← true;
else
x ← x − ρ m ∇ f (x);
until increment k;
xk ← x; k ← k + 1;
until a termination test is satisfied.
Note that the inner loop in Algorithm 9.6 is essentially the backtracking line search
algorithm—Algorithm 3.1 of Chapter 3—with a convergence criterion added to detect
whether the minimum appears to have been found to within the accuracy implied by the difference parameter k . If the gradient estimate ∇ k f is small, or if the line search fails to find a
satisfactory new iterate (indicating that the gradient approximation ∇ k f (x) is insufficiently
accurate to produce descent in f ), we decrease the difference parameter to k+1 and proceed.
A basic convergence result for Algorithm 9.6 is the following.
Theorem 9.2.
Suppose that ∇ 2 h is Lipschitz continuous, that Algorithm 9.6 generates an infinite
sequence of iterates {xk }, and that
lim
k→∞
2
k
+
η(xk ;
k)
0.
k
Suppose,
too,
that all but a finite number of inner loops in Algorithm 9.6 terminate with
∇ f (xk ) ≤ k . Then all limit points of the sequence {xk } are stationary.
k
PROOF. Using { k } ↓ 0, we have under our assumptions on inner loop termination that
∇ k f (xk ) → 0. By invoking the error bound (9.5) and noting that the right-hand side of
this expression is approaching zero, we conclude that ∇h(xk ) → 0. Hence all limit points
satisfy ∇h(x) 0, as claimed.
241
242
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
More sophisticated versions of implicit filtering methods can be derived by using
the gradient estimate ∇ k f to construct quasi-Newton approximate Hessians, and thus
generating quasi-Newton search directions instead of the negative-approximate-gradient
search direction used in Algorithm 9.6.
NOTES AND REFERENCES
A classical reference on derivative-free methods is Brent [39], which focuses primarily
on one-dimensional problems and includes discussion of roundoff errors and global minimization. Recent surveys on derivative-free methods include Wright [314], Powell [256],
Conn, Scheinberg, and Toint [76], and Kolda, Lewis, and Torczon [183].
The first model-based method for derivative-free optimization was proposed by Winfield [307]. It uses quadratic models, which are determined by the interpolation conditions
(9.7), and computes steps by solving a subproblem of the form (9.9). Practical procedures
for improving the geometry of the interpolation points were first developed by Powell in
the context of model-based methods using linear and quadratic polynomials; see [256] for
a review of this work.
Conn, Scheinberg, and Toint [75] propose and analyze model-based methods and
study the use of Newton fundamental polynomials. Methods that combine minimum change
updating and interpolation are discussed by Powell [258, 260]. Our presentation of modelbased methods in Section 9.2 is based on [76, 259, 258].
For a comprehensive discussion of pattern-search methods of the type discussed here,
we refer the reader to the review paper of Kolda, Lewis, and Torczon [183], and the references
therein.
The method of conjugate directions given in Algorithm 9.3 was proposed by Powell [239]. For a discussion on the rate of convergence of the coordinate descent method and
for more references about this method, see Luenberger [195]. For further information on
implicit filtering, see Kelley [179] and Choi and Kelley [60] and the references therein.
Software packages that implement model-based methods include COBYLA [258], DFO
[75], UOBYQA [257], WEDGE [200], and NEWUOA [260]. The earliest code is COBYLA, which
employs linear models. DFO, UOBYQA, and WEDGE use quadratic models, whereas the method
based on minimum change updating (9.19) is implemented in NEWUOA. A pattern-search
method is implemented in APPS [171], while DIRECT [173] is designed to find a global solution.
✐
EXERCISES
✐
9.1 Prove Lemma 9.1.
✐
9.2
(a) Verify that the number of interpolation conditions to uniquely determine the
coefficients in (9.6) are q 12 (n + 1)(n + 2).
9.6.
IMPLICIT FILTERING
(b) Verify that the number of vertices and midpoints of the edges of a nondegenerate
simplex in R n add up to q 12 (n + 1)(n + 2) and can therefore be used as the initial
interpolation set in a DFO algorithm.
(c) How many interpolation conditions would be required to determine the coefficients
in (9.6) if the matrix G were identically 0? How many if G were diagonal? How many
if G were tridiagonal?
✐
9.3 Describe conditions on the vectors s l that guarantee that the model (9.14) is
uniquely determined.
✐
9.4 Consider the determination of a quadratic function in two variables.
(a) Show that six points on a line do not determine the quadratic.
(b) Show that six points in a circle in the plane do not uniquely determine the quadratic.
✐
9.5 Use induction to show that at the end of the outer iteration k of Algorithm 9.3, the
directions pn−k , pn−k+1 , . . . , pn are conjugate. Use this fact to show that if the step lengths
αi in Algorithm 9.3 are never zero, the iteration terminates at the minimizer of (9.25) after
at most n outer iterations.
✐
9.6 Write a program that computes the one-dimensional minimizer of a strictly
convex quadratic function f along a direction p using quadratic interpolation. Describe
the formulas used in your program.
✐
9.7 Find the quadratic function
1
1
m(x1 , x2 ) f + g1 x1 + g2 x2 + G 211 x12 + G 12 x1 x2 + G 222 x22
2
2
that interpolates the following data: x0 y 1 (0, 0)T , y 2 (1, 0)T , y 3 (2, 0)T ,
y 4 (1, 1)T , y 5 (0, 2)T , y 6 (0, 1)T , and f (y 1 ) 1, f (y 2 ) 2.0084, f (y 3 ) 7.0091,
f (y 4 ) 1.0168, f (y 5 ) −0.9909, and f (y 6 ) −0.9916.
✐
9.8 Find the value of δ for which the coordinate generating set (9.23) satisfies the
property (9.21).
✐ 9.9 Show that κ(Dk ) 0, where κ(·) is defined by (9.21) and Dk {ei , −ei } for any
i 1, 2, . . . , n.
✐
9.10 (Hard) Prove that the generating set (9.24) satisfies the property (9.21) for a
certain value δ > 0, and find this value of δ.
✐
9.11 Justify the statement that the average function value at the Nelder–Mead simplex
points will decrease over one step if any of the points x̄(−1), x̄(−2), x̄(− 12 ), x̄( 21 ) are adopted
as a replacement for xn+1 .
243
244
CHAPTER 9.
DERIVATIVE-FREE OPTIMIZATION
✐
9.12 Show that if f is a convex function, the shrinkage step in the Nelder–Mead
simplex method will not increase the average value of the function over the simplex vertices
defined by (9.29). Show that unless f (x1 ) f (x2 ) · · · f (xn+1 ), the average value will
in fact decrease.
✐
9.13 Suppose for the f defined in (9.2), we define the approximate gradient ∇ f (x)
by the forward-difference formula
∇ f (x) f (x + ei ) − f (x)
,
i1,2,...,n
rather than the central-difference formula (9.3). (This formula requires only half as many
function evaluations but is less accurate.) For this definition, prove the following variant
of Lemma 9.1: Suppose that ∇h(x) is Lipschitz continuous in a neighborhood of the box
{z | z ≥ x, z − x∞ ≤ } with Lipschitz constant L h . Then we have
∇ f (x) − ∇h(x)∞ ≤ L h +
η(x; )
where η(x; ) is redefined as follows:
η(x; ) sup
z≥x, z−x∞ ≤
|φ(z)|.
,
This is page 245
Printer: Opaque this
10
CHAPTER
Least-Squares
Problems
In least-squares problems, the objective function f has the following special form:
f (x) 1
2
m
r 2j (x),
(10.1)
j1
where each r j is a smooth function from IRn to IR. We refer to each r j as a residual, and we
assume throughout this chapter that m ≥ n.
Least-squares problems arise in many areas of applications, and may in fact be the
largest source of unconstrained optimization problems. Many who formulate a parametrized
246
CHAPTER 10.
LEAST-SQUARES PROBLEMS
model for a chemical, physical, financial, or economic application use a function of the
form (10.1) to measure the discrepancy between the model and the observed behavior of
the system (see Example 2.1, for instance). By minimizing this function, they select values
for the parameters that best match the model to the data. In this chapter we show how to
devise efficient, robust minimization algorithms by exploiting the special structure of the
function f and its derivatives.
To see why the special form of f often makes least-squares problems easier to solve than
general unconstrained minimization problems, we first assemble the individual components
r j from (10.1) into a residual vector r : IRn → IRm , as follows
r (x) (r1 (x), r2 (x), . . . , rm (x))T .
(10.2)
Using this notation, we can rewrite f as f (x) 12 r (x)22 . The derivatives of f (x) can be
expressed in terms of the Jacobian J (x), which is the m × n matrix of first partial derivatives
of the residuals, defined by
⎡
∂r j
J (x) ∂ xi
j1,2,...,m
i1,2,...,n
∇r1 (x)T
⎢
⎢ ∇r2 (x)T
⎢
⎢
..
⎢
.
⎣
∇rm (x)T
⎤
⎥
⎥
⎥
⎥,
⎥
⎦
(10.3)
where each ∇r j (x), j 1, 2, . . . , m is the gradient of r j . The gradient and Hessian of f
can then be expressed as follows:
∇ f (x) m
r j (x)∇r j (x) J (x)T r (x),
(10.4)
j1
∇ 2 f (x) m
∇r j (x)∇r j (x)T +
j1
J (x)T J (x) +
m
r j (x)∇ 2r j (x)
j1
m
r j (x)∇ 2r j (x).
(10.5)
j1
In many applications, the first partial derivatives of the residuals and hence the
Jacobian matrix J (x) are relatively easy or inexpensive to calculate. We can thus obtain
the gradient ∇ f (x) as written in formula (10.4). Using J (x), we also can calculate the
first term J (x)T J (x) in the Hessian ∇ 2 f (x) without evaluating any second derivatives of
the functions r j . This availability of part of ∇ 2 f (x) “for free” is the distinctive feature of
least-squares problems. Moreover, this term J (x)T J (x) is often more important than the
second summation term in (10.5), either because the residuals r j are close to affine near
the solution (that is, the ∇ 2r j (x) are relatively small) or because of small residuals (that
10.1.
BACKGROUND
is, the r j (x) are relatively small). Most algorithms for nonlinear least-squares exploit these
structural properties of the Hessian.
The most popular algorithms for minimizing (10.1) fit into the line search and
trust-region frameworks described in earlier chapters. They are based on the Newton and
quasi-Newton approaches described earlier, with modifications that exploit the particular
structure of f .
Section 10.1 contains some background on applications. Section 10.2 discusses linear least-squares problems, which provide important motivation for algorithms for the
nonlinear problem. Section 10.3 describes the major algorithms, while Section 10.4 briefly
describes a variant of least squares known as orthogonal distance regression.
Throughout this chapter, we use the notation · to denote the Euclidean norm · 2 ,
unless a subscript indicates that some other norm is intended.
10.1
BACKGROUND
We discuss a simple parametrized model and show how least-squares techniques can be
used to choose the parameters that best fit the model to the observed data.
❏ EXAMPLE 10.1
We would like to study the effect of a certain medication on a patient. We draw blood
samples at certain times after the patient takes a dose, and measure the concentration of the
medication in each sample, tabulating the time t j and concentration y j for each sample.
Based on our previous experience in such experiments, we find that the following
function φ(x; t) provides a good prediction of the concentration at time t, for appropriate
values of the five-dimensional parameter vector x (x1 , x2 , x3 , x4 , x5 ):
φ(x; t) x1 + t x2 + t 2 x3 + x4 e−x5 t .
(10.6)
We choose the parameter vector x so that this model best agrees with our observation, in
some sense. A good way to measure the difference between the predicted model values and
the observations is the following least-squares function:
1
2
m
[φ(x; t j ) − y j ]2 ,
(10.7)
j1
which sums the squares of the discrepancies between predictions and observations at each
t j . This function has precisely the form (10.1) if we define
r j (x) φ(x; t j ) − y j .
(10.8)
247
248
CHAPTER 10.
LEAST-SQUARES PROBLEMS
y
t1
t2
t3
t4
t5
t6
t7
t
Figure 10.1 Model (10.7) (smooth curve) and the observed measurements, with
deviations indicated by vertical dotted lines.
Graphically, each term in (10.7) represents the square of the vertical distance between
the curve φ(x; t) (plotted as a function of t) and the point (t j , y j ), for a fixed choice
of parameter vector x; see Figure 10.1. The minimizer x ∗ of the least-squares problem is
the parameter vector for which the sum of squares of the lengths of the dotted lines in
Figure 10.1 is minimized. Having obtained x ∗ , we use φ(x ∗ ; t) to estimate the concentration
of medication remaining in the patient’s bloodstream at any time t.
❐
This model is an example of what statisticians call a fixed-regressor model. It assumes
that the times t j at which the blood samples are drawn are known to high accuracy, while
the observations y j may contain more or less random errors due to the limitations of the
equipment (or the lab technician!)
In general data-fitting problems of the type just described, the ordinate t in the model
φ(x; t) could be a vector instead of a scalar. (In the example above, for instance, t could
have two dimensions, with the first dimension representing the time since the drug was
admistered and the second dimension representing the weight of the patient. We could then
use observations for an entire population of patients, not just a single patient, to obtain the
“best” parameters for this model.)
The sum-of-squares function (10.7) is not the only way of measuring the discrepancy
between the model and the observations. Other common measures include the maximum
absolute value
max
j1,2,...,m
|φ(x; t j ) − y j |
(10.9)
10.1.
BACKGROUND
and the sum of absolute values
m
|φ(x; t j ) − y j |.
(10.10)
j1
By using the definitions of the ∞ and 1 norms, we can rewrite these two measures as
f (x) r (x)∞ ,
f (x) r (x)1 ,
(10.11)
respectively. As we discuss in Chapter 17, the problem of minimizing the functions (10.11)
can be reformulated a smooth constrained optimization problem.
In this chapter we focus only on the 2 -norm formulation (10.1). In some situations,
there are statistical motivations for choosing the least-squares criterion. Changing the notation slightly, let the discrepancies between model and observation be denoted by j , that
is,
j
φ(x; t j ) − y j .
It often is reasonable to assume that the j ’s are independent and identically distributed
with a certain variance σ 2 and probability density function gσ (·). (This assumption will
often be true, for instance, when the model accurately reflects the actual process, and when
the errors made in obtaining the measurements y j do not contain a systematic bias.) Under
this assumption, the likelihood of a particular set of observations y j , j 1, 2, . . . , m, given
that the actual parameter vector is x, is given by the function
p(y; x, σ ) m
1
gσ ( j ) j1
m
1
gσ (φ(x; t j ) − y j ).
(10.12)
j1
Given the observations y1 , y2 , . . . , ym , the “most likely” value of x is obtained by maximizing
p(y; x, σ ) with respect to x. The resulting value of x is called the maximum likelihood
estimate.
When we assume that the discrepancies follow a normal distribution, we have
1
gσ ( ) √
2π σ 2
exp −
2
2σ 2
.
Substitution in (10.12) yields
⎛
⎞
m
1
p(y; x, σ ) (2π σ 2 )−m/2 exp ⎝− 2
[φ(x; t j ) − y j ]2 ⎠ .
2σ j1
249
250
CHAPTER 10.
LEAST-SQUARES PROBLEMS
For any fixed value of the variance σ 2 , it is obvious that p is maximized when the sum
of squares (10.7) is minimized. To summarize: When the discrepancies are assumed to be
independent and identically distributed with a normal distribution function, the maximum
likelihood estimate is obtained by minimizing the sum of squares.
The assumptions on j in the previous paragraph are common, but they do not
describe the only situation for which the minimizer of the sum of squares makes good
statistical sense. Seber and Wild [280] describe many instances in which minimization of
functions like (10.7), or generalizations of this function such as
r (x)T W r (x), where W ∈ IRm×m is symmetric,
is the crucial step in obtaining estimates of the parameters x from observed data.
10.2
LINEAR LEAST-SQUARES PROBLEMS
Many models φ(x; t) in data-fitting problems are linear functions of x. In these cases, the
residuals r j (x) defined by (10.8) also are linear, and the problem of minimizing (10.7) is
called a linear least-squares problem. We can write the residual vector as r (x) J x − y for
some matrix J and vector y, both independent of x, so that the objective is
f (x) 12 J x − y2 ,
(10.13)
where y r (0). We also have
∇ f (x) J T (J x − y),
∇ 2 f (x) J T J.
(Note that the second term in ∇ 2 f (x) (see (10.5)) disappears, because ∇ 2r j 0 for all
j 1, 2, . . . , m.) It is easy to see that the f (x) in (10.13) is convex—a property that does
not necessarily hold for the nonlinear problem (10.1). Theorem 2.5 tells us that any point x ∗
for which ∇ f (x ∗ ) 0 is the global minimizer of f . Therefore, x ∗ must satisfy the following
linear system of equations:
J T J x ∗ J T y.
(10.14)
These are known as the normal equations for (10.13).
We outline briefly three major algorithms for the unconstrained linear least-squares
problem. We assume in most of our discussion that m ≥ n and that J has full column
rank.
10.2.
LINEAR LEAST-SQUARES PROBLEMS
The first and most obvious algorithm is simply to form and solve the system (10.14)
by the following three-step procedure:
• compute the coefficient matrix J T J and the right-hand-side J T y;
• compute the Cholesky factorization of the symmetric matrix J T J ;
• perform two triangular substitutions with the Cholesky factors to recover the solution
x ∗.
The Cholesky factorization
J T J R̄ T R̄
(10.15)
(where R̄ is an n × n upper triangular with positive diagonal elements) is guaranteed
to exist when m ≥ n and J has rank n. This method is frequently used in practice
and is often effective, but it has one significant disadvantage, namely, that the condition number of J T J is the square of the condition number of J . Since the relative error
in the computed solution of a problem is usually proportional to the condition number, the Cholesky-based method may result in less accurate solutions than those obtained
from methods that avoid this squaring of the condition number. When J is ill conditioned, the Cholesky factorization process may even break down, since roundoff errors
may cause small negative elements to appear on the diagonal during the factorization
process.
A second approach is based on a QR factorization of the matrix J . Since the Euclidean
norm of any vector is not affected by orthogonal transformations, we have
J x − y Q T (J x − y)
(10.16)
for any m × m orthogonal matrix Q. Suppose we perform a QR factorization with column
pivoting on the matrix J (see (A.24)) to obtain
J Q
R
0
Q1
Q2
R
0
Q 1 R,
where
is an n × n permutation matrix (hence, orthogonal);
Q is m × m orthogonal;
Q 1 is the first n columns of Q, while Q 2 contains the last m − n columns;
R is n × n upper triangular with positive diagonal elements.
(10.17)
251
252
CHAPTER 10.
LEAST-SQUARES PROBLEMS
By combining (10.16) and (10.17), we obtain
2
QT
1
T
J x − y22 x
−
y)
(J
Q 2T
2
T 2
R
Q1 y (T x) −
T
0
Q2 y 2 2
R(T x) − Q 1T y 2 + Q 2T y .
(10.18)
No choice of x has any effect on the second term of this last expression, but we can minimize
J x − y by driving the first term to zero, that is, by setting
x ∗ R −1 Q 1T y.
(In practice, we perform a triangular substitution to solve Rz Q 1T y, then permute the
components of z to obtain x ∗ z.)
This QR-based approach does not degrade the conditioning of the problem unnecessarily. The relative error in the final computed solution x ∗ is usually proportional to the
condition number of J , not its square, and this method is usually reliable. Some situations,
however, call for greater robustness or more information about the sensitivity of the solution to perturbations in the data (J or y). A third approach, based on the singular-value
decomposition (SVD) of J , can be used in these circumstances. Recall from (A.15) that the
SVD of J is given by
J U
S
0
V T
U1
U2
S
0
V T U1 SV T ,
(10.19)
where
U is m × m orthogonal;
U1 contains the first n columns of U , U2 the last m − n columns;
V is n × n orthogonal;
S is n × n diagonal, with diagonal elements σ1 ≥ σ2 ≥ · · · ≥ σn > 0.
(Note that J T J V S 2 V T , so that the columns of V are eigenvectors of J T J with
eigenvalues σ j2 , j 1, 2, . . . , n.) By following the same logic that led to (10.18), we obtain
T 2
S
U1
(V T x) −
J x − y2 y
T
0
U2
S(V T x) − U1T y2 + U2T y2 .
(10.20)
10.2.
LINEAR LEAST-SQUARES PROBLEMS
Again, the optimum is found by choosing x to make the first term equal to zero; that is,
x ∗ V S −1 U1T y.
Denoting the ith columns of U and V by u i ∈ IRm and vi ∈ IRn , respectively, we have
x∗ n
uT y
i
i1
σi
vi .
(10.21)
This formula yields useful information about the sensitivity of x ∗ . When σi is small, x ∗
is particularly sensitive to perturbations in y that affect u iT y, and also to perturbations in
J that affect this same quantity. Such information is particularly useful when J is nearly
rank-deficient, that is, when σn /σ1 1. It is sometimes worth the extra cost of the SVD
algorithm to obtain this sensitivity information.
All three approaches above have their place. The Cholesky-based algorithm is particularly useful when m n and it is practical to store J T J but not J itself. It can also be less
expensive than the alternatives when m n and J is sparse. However, this approach must
be modified when J is rank-deficient or ill conditioned to allow pivoting of the diagonal
elements of J T J . The QR approach avoids squaring of the condition number and hence
may be more numerically robust. While potentially the most expensive, the SVD approach
is the most robust and reliable of all. When J is actually rank-deficient, some of the singular
values σi are exactly zero, and any vector x ∗ of the form
x∗ uT y
i
vi +
τi vi
σi
σ 0
σ 0
i
(10.22)
i
(for arbitrary coefficients τi ) is a minimizer of (10.20). Frequently, the solution with smallest
norm is the most desirable, and we obtain it by setting each τi 0 in (10.22). When J has
full rank but is ill conditioned, the last few singular values σn , σn−1 , . . . are small relative
to σ1 . The coefficients u iT y/σi in (10.22) are particularly sensitive to perturbations in u iT y
when σi is small, so an approximate solution that is less sentitive to perturbations than the
true solution can be obtained by omitting these terms from the summation.
When the problem is very large, it may be efficient to use iterative techniques, such
as the conjugate gradient method, to solve the normal equations (10.14). A direct implementation of conjugate gradients (Algorithm 5.2) requires one matrix vector multiplication
with J T J to be performed at each iteration. This operation can be performed by means of
successive multiplications by J and J T ; we need only the ability to perform matrix-vector
multiplications with these two matrices to implement this algorithm. Several modifications
of the conjugate gradient approach have been proposed that involve a similar amount of
work per iteration (one matrix-vector multiplication each with J and J T ) but that have
superior numerical properties. Some alternatives are described by Paige and Saunders [234],
253
254
CHAPTER 10.
LEAST-SQUARES PROBLEMS
who propose in particular an algorithm called LSQR which has become the basis of a highly
successful code.
10.3
ALGORITHMS FOR NONLINEAR LEAST-SQUARES
PROBLEMS
THE GAUSS–NEWTON METHOD
We now describe methods for minimizing the nonlinear objective function (10.1) that
exploit the structure in the gradient ∇ f (10.4) and Hessian ∇ 2 f (10.5). The simplest of these
methods—the Gauss–Newton method—can be viewed as a modified Newton’s method with
line search. Instead of solving the standard Newton equations ∇ 2 f (xk ) p −∇ f (xk ), we
solve instead the following system to obtain the search direction pkGN :
JkT Jk pkGN −JkT rk .
(10.23)
This simple modification gives a number of advantages over the plain Newton’s method.
First, our use of the approximation
∇ 2 f k ≈ JkT Jk
(10.24)
saves us the trouble of computing the individual residual Hessians ∇ 2r j , j 1, 2, . . . , m,
which are needed in the second term in (10.5). In fact, if we calculated the Jacobian Jk in the
course of evaluating the gradient ∇ f k JkT rk , the approximation (10.24) does not require
any additional derivative evaluations, and the savings in computational time can be quite
significant in some applications. Second, there are many interesting situations in which the
first term J T J in (10.5) dominates the second term (at least close to the solution x ∗ ), so
that JkT Jk is a close approximation to ∇ 2 f k and the convergence rate of Gauss–Newton is
similar to that of Newton’s method. The first term in (10.5) will be dominant when the
norm of each second-order term (that is, |r j (x)|∇ 2r j (x)) is significantly smaller than the
eigenvalues of J T J . As mentioned in the introduction, we tend to see this behavior when
either the residuals r j are small or when they are nearly affine (so that the ∇ 2r j are small).
In practice, many least-squares problems have small residuals at the solution, leading to
rapid local convergence of Gauss–Newton.
A third advantage of Gauss–Newton is that whenever Jk has full rank and the gradient
∇ f k is nonzero, the direction pkGN is a descent direction for f , and therefore a suitable
direction for a line search. From (10.4) and (10.23) we have
( pkGN )T ∇ f k ( pkGN )T JkT rk −( pkGN )T JkT Jk pkGN −Jk pkGN 2 ≤ 0.
(10.25)
10.3.
ALGORITHMS
FOR
NONLINEAR LEAST-SQUARES PROBLEMS
The final inequality is strict unless Jk pkGN 0, in which case we have by (10.23) and full rank
of Jk that JkT rk ∇ f k 0; that is, xk is a stationary point. Finally, the fourth advantage
of Gauss–Newton arises from the similarity between the equations (10.23) and the normal
equations (10.14) for the linear least-squares problem. This connection tells us that pkGN is
in fact the solution of the linear least-squares problem
min 21 Jk p + rk 2 .
(10.26)
p
Hence, we can find the search direction by applying linear least-squares algorithms to the
subproblem (10.26). In fact, if the QR or SVD-based algorithms are used, there is no need
to calculate the Hessian approximation JkT Jk in (10.23) explicitly; we can work directly with
the Jacobian Jk . The same is true if we use a conjugate-gradient technique to solve (10.26).
For this method we need to perform matrix-vector multiplications with JkT Jk , which can
be done by first multiplying by Jk and then by JkT .
If the number of residuals m is large while the number of variables n is relatively
small, it may be unwise to store the Jacobian J explicitly. A preferable strategy may be to
calculate the matrix J T J and gradient vector J T r by evaluating r j and ∇r j successively for
j 1, 2, . . . , m and performing the accumulations
JT J m
i1
(∇r j )(∇r j )T ,
JTr m
r j (∇r j ).
(10.27)
i1
The Gauss–Newton steps can then be computed by solving the system (10.23) of normal
equations directly.
The subproblem (10.26) suggests another motivation for the Gauss–Newton search
direction. We can view this equation as being obtained from a linear model for the the vector
function r (xk + p) ≈ rk + Jk p, substituted into the function 12 · 2 . In other words, we
use the approximation
f (xk + p) 12 r (xk + p)2 ≈ 12 Jk p + rk 2 ,
and choose pkGN to be the minimizer of this approximation.
Implementations of the Gauss–Newton method usually perform a line search in the
direction pkGN , requiring the step length αk to satisfy conditions like those discussed in
Chapter 3, such as the Armijo and Wolfe conditions; see (3.4) and (3.6).
CONVERGENCE OF THE GAUSS–NEWTON METHOD
The theory of Chapter 3 can applied to study the convergence properties of the
Gauss–Newton method. We prove a global convergence result under the assumption that
the Jacobians J (x) have their singular values uniformly bounded away from zero in the
255
256
CHAPTER 10.
LEAST-SQUARES PROBLEMS
region of interest; that is, there is a constant γ > 0 such that
J (x)z ≥ γ z
(10.28)
for all x in a neighborhood N of the level set
L {x | f (x) ≤ f (x0 )},
(10.29)
where x0 is the starting point for the algorithm. We assume here and in the rest of the chapter
that L is bounded. Our result is a consequence of Theorem 3.2.
Theorem 10.1.
Suppose each residual function r j is Lipschitz continuously differentiable in a neighborhood N of the bounded level set (10.29), and that the Jacobians J (x) satisfy the uniform
full-rank condition (10.28) on N . Then if the iterates xk are generated by the Gauss–Newton
method with step lengths αk that satisfy (3.6), we have
lim JkT rk 0.
k→∞
PROOF. First, we note that the neighborhood N of the bounded level set L can be chosen
small enough that the following properties are satisfied for some positive constants L and β:
|r j (x)| ≤ β and ∇r j (x) ≤ β,
|r j (x) − r j (x̃)| ≤ Lx − x̃ and ∇r j (x) − ∇r j (x̃) ≤ Lx − x̃,
for all x, x̃ ∈ N and all j 1, 2, . . . , m. It is easy to deduce that there exists a constant
β̄ > 0 such that J (x)T J (x) ≤ β̄ for all x ∈ L. In addition, by applying the
results concerning Lipschitz continuity of products and sums (see for example (A.43)) to
the gradient ∇ f (x) mj1 r j (x)∇r j (x), we can show that ∇ f is Lipschitz continuous.
Hence, the assumptions of Theorem 3.2 are satisfied.
We check next that the angle θk between the search direction pkGN and the negative
gradient −∇ f k is uniformly bounded away from π/2. From (3.12), (10.25), and (10.28),
we have for x xk ∈ L and p GN pkGN that
cos θk −
J p GN 2
γ 2 p GN 2
(∇ f )T p GN
γ2
≥ 2 GN 2 2 > 0.
GN
GN
GN
T
p ∇ f p J J p β̄ p β̄
It follows from (3.14) in Theorem 3.2 that ∇ f (xk ) → 0, giving the result.
If Jk is rank-deficient for some k (so that a condition like (10.28) is not satisfied), the
coefficient matrix in (10.23) is singular. The system (10.23) still has a solution, however,
because of the equivalence between this linear system and the minimization problem (10.26).
10.3.
ALGORITHMS
NONLINEAR LEAST-SQUARES PROBLEMS
FOR
In fact, there are infinitely many solutions for pkGN in this case; each of them has the form
of (10.22). However, there is no longer an assurance that cos θk is uniformly bounded away
from zero, so we cannot prove a result like Theorem 10.1.
The convergence of Gauss–Newton to a solution x ∗ can be rapid if the leading term
T
Jk Jk dominates the second-order term in the Hessian (10.5). Suppose that xk is close to
x ∗ and that assumption (10.28) is satisfied. Then, applying an argument like the Newton’s
method analysis (3.31), (3.32), (3.33) in Chapter 3, we have for a unit step in the Gauss–
Newton direction that
xk + pkGN − x ∗ xk − x ∗ − [J T J (xk )]−1 ∇ f (xk )
[J T J (xk )]−1 J T J (xk )(xk − x ∗ ) + ∇ f (x ∗ ) − ∇ f (xk ) ,
where J T J (x) is shorthand notation for J (x)T J (x). Using H (x) to denote the second-order
term in (10.5), we have from (A.57) that
∇ f (xk ) − ∇ f (x ∗ ) 1
0
J T J (x ∗ + t(xk − x ∗ ))(xk − x ∗ ) dt
1
+
H (x ∗ + t(xk − x ∗ ))(xk − x ∗ ) dt.
0
A similar argument as in (3.32), (3.33), assuming Lipschitz continuity of H (·) near x ∗ ,
shows that
xk + pkGN − x ∗ 1
[J T J (xk )]−1 H (x ∗ + t(xk − x ∗ ))xk − x ∗ dt + O(xk − x ∗ 2 )
≤
0
≈ [J T J (x ∗ )]−1 H (x ∗ ) xk − x ∗ + O(xk − x ∗ 2 ).
(10.30)
Hence, if [J T J (x ∗ )]−1 H (x ∗ ) 1, we can expect a unit step of Gauss–Newton to move
us much closer to the solution x ∗ , giving rapid local convergence. When H (x ∗ ) 0, the
convergence is actually quadratic.
When n and m are both large and the Jacobian J (x) is sparse, the cost of computing
steps exactly by factoring either Jk or JkT Jk at each iteration may become quite expensive
relative to the cost of function and gradient evaluations. In this case, we can design inexact
variants of the Gauss–Newton algorithm that are analogous to the inexact Newton algorithms discussed in Chapter 7. We simply replace the Hessian ∇ 2 f (xk ) in these methods by
its approximation JkT Jk . The positive semidefiniteness of this approximation simplifies the
resulting algorithms in several places.
257
258
CHAPTER 10.
LEAST-SQUARES PROBLEMS
THE LEVENBERG–MARQUARDT METHOD
Recall that the Gauss–Newton method is like Newton’s method with line search, except
that we use the convenient and often effective approximation (10.24) for the Hessian. The
Levenberg–Marquardt method can be obtained by using the same Hessian approximation,
but replacing the line search with a trust-region strategy. The use of a trust region avoids
one of the weaknesses of Gauss–Newton, namely, its behavior when the Jacobian J (x) is
rank-deficient, or nearly so. Since the same Hessian approximations are used in each case,
the local convergence properties of the two methods are similar.
The Levenberg–Marquardt method can be described and analyzed using the trustregion framework of Chapter 4. (In fact, the Levenberg–Marquardt method is sometimes
considered to be the progenitor of the trust-region approach for general unconstrained
optimization discussed in Chapter 4.) For a spherical trust region, the subproblem to be
solved at each iteration is
min 12 Jk p + rk 2 ,
subject to p ≤ k ,
p
(10.31)
where k > 0 is the trust-region radius. In effect, we are choosing the model function m k (·)
in (4.3) to be
m k ( p) 12 rk 2 + p T JkT rk + 12 p T JkT Jk p.
(10.32)
We drop the iteration counter k during the rest of this section and concern ourselves
with the subproblem (10.31). The results of Chapter 4 allow us to characterize the solution
of (10.31) in the following way: When the solution p GN of the Gauss–Newton equations
(10.23) lies strictly inside the trust region (that is, p GN < ), then this step p GN also solves
the subproblem (10.31). Otherwise, there is a λ > 0 such that the solution p p LM of
(10.31) satisfies p and
J T J + λI p −J T r.
(10.33)
This claim is verified in the following lemma, which is a straightforward consequence of
Theorem 4.1 from Chapter 4.
Lemma 10.2.
The vector p LM is a solution of the trust-region subproblem
min J p + r 2 ,
subject to p ≤ ,
p
if and only if p LM is feasible and there is a scalar λ ≥ 0 such that
(J T J + λI ) p LM −J T r,
λ( − p ) 0.
LM
(10.34a)
(10.34b)
10.3.
ALGORITHMS
FOR
NONLINEAR LEAST-SQUARES PROBLEMS
PROOF. In Theorem 4.1, the semidefiniteness condition (4.8c) is satisfied automatically,
since J T J is positive semidefinite and λ ≥ 0. The two conditions (10.34a) and (10.34b)
follow from (4.8a) and (4.8b), respectively.
Note that the equations (10.33) are just the normal equations for the following linear
least-squares problem:
min
p
1
2
2
J
r p+
√
.
0 λI
(10.35)
Just as in the Gauss–Newton case, the equivalence between (10.33) and (10.35) gives us a
way of solving the subproblem without computing the matrix–matrix product J T J and its
Cholesky factorization.
IMPLEMENTATION OF THE LEVENBERG–MARQUARDT METHOD
To find a value of λ that approximately matches the given in Lemma 10.2, we can
use the rootfinding algorithm described in Chapter 4. It is easy to safeguard this procedure:
The Cholesky factor R is guaranteed to exist whenever the current estimate λ() is positive,
since the approximate Hessian B J T J is already positive semidefinite. Because of the
special structure of B, we do not need to compute the Cholesky factorization of B + λI
from scratch in each iteration of Algorithm 4.1. Rather, we present an efficient technique
for finding the following QR factorization of the coefficient matrix in (10.35):
Rλ
0
Q λT
J
√
λI
(10.36)
(Q λ orthogonal, Rλ upper triangular). The upper triangular factor Rλ satisfies RλT Rλ (J T J + λI ).
We can save computer time in the calculation of the factorization (10.36) by using
a combination of Householder and Givens transformations. Suppose we use Householder
transformations to calculate the QR factorization of J alone as
JQ
R
.
0
(10.37)
We then have
⎡
⎢
⎣
R
0
√
λI
⎤
⎥
⎦
QT
I
J
√
λI
.
(10.38)
259
260
CHAPTER 10.
LEAST-SQUARES PROBLEMS
The leftmost matrix in this formula is upper triangular except for the n nonzero terms of
the matrix λI . These can be eliminated by a sequence of n(n + 1)/2 Givens rotations, in
which the diagonal elements of the upper triangular part are used to eliminate the nonzeros
of λI and the fill-in terms that arise in the process. The first few steps of this process are as
follows:
√
√
rotate row n of R with row n of λI , to eliminate the (n, n) element of λI ;
√
rotate row n − 1 of R with row n − 1 of λI to eliminate the (n − 1, n −√1) element
of the latter matrix. This step introduces fill-in in position
√ (n − 1, n) of λI , which
is eliminated by rotating row n of R with row n − 1 of λI , to eliminate the fill-in
element at position (n − 1, n);
√
rotate row n − 2 of R with row n − 2 of λI , to eliminate the (n − 2) diagonal in the
latter matrix. This step introduces fill-in in the (n − 2, n − 1) and (n − 2, n) positions,
which we eliminate by · · ·
and so on. If we gather all the Givens rotations into a matrix Q̄ λ , we obtain from (10.38)
that
⎡
⎢
Q̄ λT ⎣
R
⎤
⎡
Rλ
⎤
⎥ ⎢
⎥
0 ⎦ ⎣ 0 ⎦,
√
0
λI
and hence (10.36) holds with
Qλ Q
I
Q̄ λ .
The advantage of this combined approach is that when the value of λ is changed in the
rootfinding algorithm, we need only recalculate Q̄ λ and not the Householder part of the
factorization (10.38). This feature can save a lot of computation in the case of m n, since
just O(n 3 ) operations are required to recalculate Q̄ λ and Rλ for each value of λ, after the
initial cost of O(mn 2 ) operations needed to calculate Q in (10.37).
Least-squares problems are often poorly scaled. Some of the variables could have
values of about 104 , while other variables could be of order 10−6 . If such wide variations are
ignored, the algorithms above may encounter numerical difficulties or produce solutions of
poor quality. One way to reduce the effects of poor scaling is to use an ellipsoidal trust region
in place of the spherical trust region defined above. The step is confined to an ellipse in
which the lengths of the principal axes are related to the typical values of the corresponding
variables. Analytically, the trust-region subproblem becomes
min 12 Jk p + rk 2 ,
p
subject to Dk p ≤ k ,
(10.39)
10.3.
ALGORITHMS
FOR
NONLINEAR LEAST-SQUARES PROBLEMS
where Dk is a diagonal matrix with positive diagonal entries (cf. (7.13)). Instead of (10.33),
the solution of (10.39) satisfies an equation of the form
JkT Jk + λDk2 pkLM −JkT rk ,
(10.40)
and, equivalently, solves the linear least-squares problem
2
Jk
rk min √
p+
.
p 0 λDk
(10.41)
The diagonals of the scaling matrix Dk can change from iteration to iteration, as we gather
information about the typical range of values for each component of x. If the variation in
these elements is kept within certain bounds, then the convergence theory for the spherical
case continues to hold, with minor modifications. Moreover, the technique described above
for calculating Rλ needs no modification. Seber and Wild [280] suggest choosing the
diagonals of Dk2 to match those of JkT Jk , to make the algorithm invariant under diagonal
scaling of the components of x. This approach is analogous to the technique of scaling
by diagonal elements of the Hessian, which was described in Section 4.5 in the context of
trust-region algorithms for unconstrained optimization.
For problems in which m and n are large and J (x) is sparse, we may prefer to solve
(10.31) or (10.39) approximately using the CG-Steihaug algorithm, Algorithm 7.2 from
Chapter 7, with JkT Jk replacing the exact Hessian ∇ 2 f k . Positive semidefiniteness of the
matrix JkT Jk makes for some simplification of this algorithm, because negative curvature
cannot arise. It is not necessary to calculate JkT Jk explicitly to implement Algorithm 7.2; the
matrix-vector products required by the algorithm can be found by forming matrix-vector
products with Jk and JkT separately.
CONVERGENCE OF THE LEVENBERG–MARQUARDT METHOD
It is not necessary to solve the trust-region problem (10.31) exactly in order for
the Levenberg–Marquardt method to enjoy global convergence properties. The following
convergence result can be obtained as a direct consequence of Theorem 4.6.
Theorem 10.3.
Let η ∈ 0, 14 in Algorithm 4.1 of Chapter 4, and suppose that the level set L defined
in (10.29) is bounded and that the residual functions r j (·), j 1, 2, . . . , m are Lipschitz
continuously differentiable in a neighborhood N of L. Assume that for each k, the approximate
solution pk of (10.31) satisfies the inequality
J T rk m k (0) − m k ( pk ) ≥ c1 JkT rk min k , kT
,
Jk Jk (10.42)
261
262
CHAPTER 10.
LEAST-SQUARES PROBLEMS
for some constant c1 > 0, and in addition pk ≤ γ k for some constant γ ≥ 1. We then
have that
lim ∇ f k lim JkT rk 0.
k→∞
k→∞
PROOF. The smoothness assumption on r j (·) implies that we can choose a constant M > 0
such that JkT Jk ≤ M for all iterates k. Note too that the objective f is bounded below
(by zero). Hence, the assumptions of Theorem 4.6 are satisfied, and the result follows
immediately.
As in Chapter 4, there is no need to calculate the right-hand-side in the inequality
(10.42) or to check it explicitly. Instead, we can simply require the decrease given by our
approximate solution pk of (10.31) to at least match the decrease given by the Cauchy point,
which can be calculated inexpensively in the same way as in Chapter 4. If we use the iterative
CG-Steihaug approach, Algorithm 7.2, the condition (10.42) is satisfied automatically for
c1 1/2, since the Cauchy point is the first estimate of pk computed by this approach,
while subsequent estimates give smaller values for the model function.
The local convergence behavior of Levenberg–Marquardt is similar to the Gauss–
Newton method. Near a solution x ∗ at which the first term of the Hessian ∇ 2 f (x ∗ ) (10.5)
dominates the second term, the model function in (10.31), the trust region becomes inactive
and the algorithm takes Gauss–Newton steps, giving the rapid local convergence expression
(10.30).
METHODS FOR LARGE-RESIDUAL PROBLEMS
In large-residual problems, the quadratic model in (10.31) is an inadequate representation of the function f because the second-order part of the Hessian ∇ 2 f (x) is too
significant to be ignored. In data-fitting problems, the presence of large residuals may
indicate that the model is inadequate or that errors have been made in monitoring the
observations. Still, the practitioner may need to solve the least-squares problem with the
current model and data, to indicate where improvements are needed in the weighting of
observations, modeling, or data collection.
On large-residual problems, the asymptotic convergence rate of Gauss–Newton and
Levenberg–Marquardt algorithms is only linear—slower than the superlinear convergence
rate attained by algorithms for general unconstrained problems, such as Newton or quasiNewton. If the individual Hessians ∇ 2r j are easy to calculate, it may be better to ignore the
structure of the least-squares objective and apply Newton’s method with trust region or line
search to the problem of minimizing f . Quasi-Newton methods, which attain a superlinear convergence rate without requiring calculation of ∇ 2r j , are another option. However,
the behavior of both Newton and quasi-Newton on early iterations (before reaching a
neighborhood of the solution) may be inferior to Gauss–Newton and Levenberg–Marquardt.
10.3.
ALGORITHMS
FOR
NONLINEAR LEAST-SQUARES PROBLEMS
Of course, we often do not know beforehand whether a problem will turn out to
have small or large residuals at the solution. It seems reasonable, therefore, to consider
hybrid algorithms, which would behave like Gauss–Newton or Levenberg–Marquardt if the
residuals turn out to be small (and hence take advantage of the cost savings associated with
these methods) but switch to Newton or quasi-Newton steps if the residuals at the solution
appear to be large.
There are a couple of ways to construct hybrid algorithms. One approach, due to
Fletcher and Xu (see Fletcher [101]), maintains a sequence of positive definite Hessian approximations Bk . If the Gauss–Newton step from xk reduces the function f by a certain
fixed amount (say, a factor of 5), then this step is taken and Bk is overwritten by JkT Jk .
Otherwise, a direction is computed using Bk , and the new point xk+1 is obtained by performing a line search. In either case, a BFGS-like update is applied to Bk to obtain a new
approximation Bk+1 . In the zero-residual case, the method eventually always takes Gauss–
Newton steps (giving quadratic convergence), while it eventually reduces to BFGS in the
nonzero-residual case (giving superlinear convergence). Numerical results in Fletcher [101,
Tables 6.1.2, 6.1.3] show good results for this approach on small-, large-, and zero-residual
problems.
A second way to combine Gauss–Newton and quasi-Newton ideas is to maintain
approximations to just the second-order part of the Hessian. That is, we maintain a sequence
of matrices Sk that approximate the summation term mj1 r j (xk )∇ 2r j (xk ) in (10.5), and
then use the overall Hessian approximation
Bk JkT Jk + Sk
in a trust-region or line search model for calculating the step pk . Updates to Sk are devised
so that the approximate Hessian Bk , or its constituent parts, mimics the behavior of the
corresponding exact quantities over the step just taken. The update formula is based on a
secant equation, which arises also in the context of unconstrained minimization (6.6) and
nonlinear equations (11.27). In the present instance, there are a number of different ways
to define the secant equation and to specify the other conditions needed for a complete
update formula for Sk . We describe the algorithm of Dennis, Gay, and Welsch [90], which
is probably the best-known algorithm in this class because of its implementation in the
well-known NL2SOL package.
In [90], the secant equation is motivated in the following way. Ideally, Sk+1 should be
a close approximation to the exact second-order term at x xk+1 ; that is,
Sk+1 ≈
m
r j (xk+1 )∇ 2r j (xk+1 ).
j1
Since we do not want to calculate the individual Hessians ∇ 2r j in this formula, we could
replace each of them with an approximation (B j )k+1 and impose the condition that (B j )k+1
263
264
CHAPTER 10.
LEAST-SQUARES PROBLEMS
should mimic the behavior of its exact counterpart ∇ 2r j over the step just taken; that
is,
(B j )k+1 (xk+1 − xk ) ∇r j (xk+1 ) − ∇r j (xk )
(row j of J (xk+1 ))T − (row j of J (xk ))T .
This condition leads to a secant equation on Sk+1 , namely,
Sk+1 (xk+1 − xk ) m
r j (xk+1 )(B j )k+1 (xk+1 − xk )
j1
m
r j (xk+1 ) (row j of J (xk+1 ))T − (row j of J (xk ))T
j1
T
Jk+1
rk+1 − JkT rk+1 .
As usual, this condition does not completely specify the new approximation Sk+1 . Dennis,
Gay, and Welsch add requirements that Sk+1 be symmetric and that the difference Sk+1 − Sk
from the previous estimate Sk be minimized in a certain sense, and derive the following
update formula:
Sk+1 Sk +
(y − Sk s)T s T
(y − Sk s)y T + y(y − Sk s)T
−
yy ,
yT s
(y T s)2
(10.43)
where
s xk+1 − xk ,
T
rk+1 − JkT rk ,
y Jk+1
T
y Jk+1
rk+1 − JkT rk+1 .
Note that (10.43) is a slight variant on the DFP update for unconstrained minimization. It
would be identical if y and y were the same.
Dennis, Gay, and Welsch use their approximate Hessian JkT Jk + Sk in conjunction
with a trust-region strategy, but a few more features are needed to enhance its performance.
One deficiency of its basic update strategy for Sk is that this matrix is not guaranteed to
vanish as the iterates approach a zero-residual solution, so it can interfere with superlinear
convergence. This problem is avoided by scaling Sk prior to its update; we replace Sk by τk Sk
on the right-hand-side of (10.43), where
|s T y |
.
τk min 1, T
|s Sk s|
10.4.
ORTHOGONAL DISTANCE REGRESSION
A final modification in the overall algorithm is that the Sk term is omitted from the Hessian
approximation when the resulting Gauss–Newton model produces a sufficiently good step.
10.4
ORTHOGONAL DISTANCE REGRESSION
In Example 10.1 we assumed that no errors were made in noting the time at which the blood
samples were drawn, so that the differences between the model φ(x; t j ) and the observation
y j were due to inadequacy in the model or measurement errors in y j . We assumed that
any errors in the ordinates—the times t j —are tiny by comparison with the errors in the
observations. This assumption often is reasonable, but there are cases where the answer can
be seriously distorted if we fail to take possible errors in the ordinates into account. Models
that take these errors into account are known in the statistics literature as errors-in-variables
models [280, Chapter 10], and the resulting optimization problems are referred to as total
least squares in the case of a linear model (see Golub and Van Loan [136, Chapter 5]) or as
orthogonal distance regression in the nonlinear case (see Boggs, Byrd, and Schnabel [30]).
We formulate this problem mathematically by introducing perturbations δ j for the
ordinates t j , as well as perturbations j for y j , and seeking the values of these 2m perturbations that minimize the discrepancy between the model and the observations, as measured
by a weighted least-squares objective function. To be precise, we relate the quantities t j , y j ,
δ j , and j by
y j φ(x; t j + δ j ) +
j,
j 1, 2, . . . , m,
(10.44)
subject to (10.44).
(10.45)
and define the minimization problem as
min
x,δ j ,
j
1
2
m
w 2j
2
j
+ d 2j δ 2j ,
j1
The quantities wi and di are weights, selected either by the modeler or by some automatic
estimate of the relative significance of the error terms.
It is easy to see how the term “orthogonal distance regression” originates when we
graph this problem; see Figure 10.2. If all the weights wi and di are equal, then each term
in the summation (10.45) is simply the shortest distance between the point (t j , y j ) and the
curve φ(x; t) (plotted as a function of t). The shortest path between each point and the
curve is orthogonal to the curve at the point of intersection.
Using the constraints (10.44) to eliminate the variables j from (10.45), we obtain the
unconstrained least-squares problem
min F(x, δ) x,δ
1
2
m
j1
w 2j [y j − φ(x; t j + δ j )]2 + d 2j δ 2j 1
2
2m
j1
r 2j (x, δ),
(10.46)
265
266
CHAPTER 10.
LEAST-SQUARES PROBLEMS
y
t1
t2
t3
t4
t5
t6
t7
t
Figure 10.2 Orthogonal distance regression minimizes the sum of squares of the
distance from each point to the curve.
where δ (δ1 , δ2 , . . . , δm )T and we have defined
r j (x, δ) w j [φ(x; t j + δ j ) − y j ],
j 1, 2, . . . , m,
d j−m δ j−m ,
j m + 1, . . . , 2m.
(10.47)
Note that (10.46) is now a standard least-squares problem with 2m residuals and m + n
unknowns, which we can solve by using the techniques in this chapter. A naive implementation of this strategy may, however, be quite expensive, since the number of parameters (2n)
and the number of observations (m + n) may both be much larger than for the original
problem.
Fortunately, the Jacobian matrix for (10.46) has a special structure that can be exploited in implementing the Gauss–Newton or Levenberg–Marquardt methods. Many of its
components are zero; for instance, we have
∂[φ(t j + δ j ; x) − y j ]
∂r j
0,
∂δi
∂δi
i, j 1, 2, . . . , m, i j,
and
∂r j
0,
∂ xi
j m + 1, . . . , 2m, i 1, 2, . . . , n.
10.4.
ORTHOGONAL DISTANCE REGRESSION
Additionally, we have for j 1, 2, . . . , m and i 1, 2, . . . , m that
∂rm+ j
∂δi
dj
if i j,
0
otherwise.
Hence, we can partition the Jacobian of the residual function r defined by (10.47) into
blocks and write
J (x, δ) Jˆ
V
0
D
,
(10.48)
where V and D are m × m diagonal matrices and Jˆ is the m × n matrix of partial derivatives
of the functions w j φ(t j + δ j ; x) with respect to x. Boggs, Byrd, and Schnabel [30]) apply
the Levenberg–Marquardt algorithm to (10.46) and note that block elimination can be used
to solve the subproblems (10.33), (10.35) efficiently. Given the partitioning (10.48), we can
partition the step vector p and the residual vector r accordingly as
p
px
pδ
,
r
r̂1
r̂2
,
and write the normal equations (10.33) in the partitioned form
JˆT Jˆ + λI
V Jˆ
JˆT V
V 2 + D 2 + λI
px
pδ
−
JˆT r̂1
V r̂1 + Dr̂2
.
(10.49)
Since the lower right submatrix V 2 + D 2 + λI is diagonal, it is easy to eliminate pδ from
this system and obtain a smaller n × n system to be solved for px alone. The total cost
of finding a step is only marginally greater than for the m × n problem arising from the
standard least-squares model.
NOTES AND REFERENCES
Algorithms for linear least squares are discussed comprehensively by Björck [29],
who includes detailed error analyses of the different algorithms and software listings. He
considers not just the basic problem (10.13) but also the situation in which there are bounds
(for example, x ≥ 0) or linear constraints (for example, Ax ≥ b) on the variables. Golub
and Van Loan [136, Chapter 5] survey the state of the art, including discussion of the
suitability of the different approaches (for example, normal equations vs. QR factorization)
for different problem types. A classical reference on linear least-squares is Lawson and
Hanson [188].
267
268
CHAPTER 10.
LEAST-SQUARES PROBLEMS
Very large nonlinear least-squares problems arise in numerous areas of application,
such as medical imaging, geophysics, economics, and engineering design. In many instances,
both the number of variables n and the number of residuals m is large, but it is also quite
common that only m is large.
The original description of the Levenberg–Marquardt algorithm [190, 203] did not
make the connection with the trust-region concept. Rather, it adjusted the value of λ in
(10.33) directly, increasing or decreasing it by a certain factor according to whether or not
the previous trial step was effective in decreasing f (·). (The heuristics for adjusting λ were
analogous to those used for adjusting the trust-region radius k in Algorithm 4.1.) Similar
convergence results to Theorem 10.3 can be proved for algorithms that use this approach
(see, for instance, Osborne [231]), independently of trust-region analysis. The connection
with trust regions was firmly established by Moré [210].
Wright and Holt [318] present an inexact Levenberg–Marquardt approach for
large-scale nonlinear least squares that manipulates the parameter λ directly rather than
making use of the connection to trust-region algorithms. This method takes steps p̄k that,
analogously to (7.2) and (7.3) in Chapter 7, satisfy the system
T
J Jk + λk I p̄k + J T rk ≤ ηk J T rk ,
k
k
k
for some ηk ∈ [0, η],
where η ∈ (0, 1) is a constant and {ηk } is a forcing sequence. A ratio of actual to predicted decrease is used to decide whether the step p̄k should be taken, and convergence
to stationary points can be proved under certain assumptions. The method can be implemented efficiently by using Algorithm LSQR of Paige and Saunders [234] to calculate the
approximate solution of (10.35) since, for a small marginal cost, this algorithm can compute
approximate solutions for a number of different values of λk simultaneously. Hence, we can
compute values of p̄k corresponding to a range of values of λk , and choose the actual step to
be the one corresponding to the smallest λk for which the actual-predicted decrease ratio is
satisfactory.
Nonlinear least squares software is fairly prevalent because of the high demand
for it. Major numerical software libraries such as IMSL, HSL, NAG, and SAS, as well as
programming environments such as Mathematica and Matlab, contain robust nonlinear
least-squares implementations. Other high quality implmentations include DFNLP, MINPACK,
NL2SOL, and NLSSOL; see Moré and Wright [217, Chapter 3]. The nonlinear programming
packages LANCELOT, KNITRO, and SNOPT provide large-scale implementions of the Gauss–
Newton and Levenberg–Marquardt methods. The orthogonal distance regression algorithm
is implemented by ORDPACK [31].
All these routines (which can be accessed through the web) give the user the option
of either supplying Jacobians explicitly or else allowing the code to compute them by finite
differencing. (In the latter case, the user need only write code to compute the residual vector
r (x); see Chapter 8.) Seber and Wild [280, Chapter 15] describe some of the important
practical issues in selecting software for statistical applications.
10.4.
ORTHOGONAL DISTANCE REGRESSION
✐
EXERCISES
✐
10.1 Let J be an m × n matrix with m ≥ n, and let y ∈ IRm be a vector.
(a) Show that J has full column rank if and only if J T J is nonsingular.
(b) Show that J has full column rank if and only if J T J is positive definite.
✐
10.2 Show that the function f (x) in (10.13) is convex.
✐
10.3 Show that
(a) if Q is an orthogonal matrix, then Qx x for any vector x;
(b) the matrices R̄ in (10.15) and R in (10.17) are identical if I , provided that J has
full column rank n.
✐
10.4
(a) Show that x ∗ defined in (10.22) is a minimizer of (10.13).
(b) Find x ∗ and conclude that this norm is minimized when τi 0 for all i with σi 0.
✐
10.5 Suppose that each residual function r j and its gradient are Lipschitz continuous
with Lipschitz constant L, that is,
r j (x) − r j (x̃) ≤ Lx − x̃,
∇r j (x) − ∇r j (x̃) ≤ Lx − x̃
for all j 1, 2, . . . , m and all x, x̃ ∈ D, where D is a compact subset of IRn . Assume also
that the r j are bounded on D, that is, there exists M > 0 such that |r j (x)| ≤ M for all
j 1, 2, . . . , m and all x ∈ D. Find Lipschitz constants for the Jacobian J (10.3) and the
gradient ∇ f (10.4) over D.
✐
10.6 Express the solution p of (10.33) in terms of the singular-value decomposition
of J (x) and the scalar λ. Express its squared-norm p2 in these same terms, and show that
lim p λ→0
uT r
i
vi .
σ
i
σ 0
i
269
This is pag
Printer: O
11
CHAPTER
Nonlinear
Equations
In many applications we do not need to optimize an objective function explicitly, but rather
to find values of the variables in a model that satisfy a number of given relationships. When
these relationships take the form of n equalities—the same number of equality conditions
as variables in the model—the problem is one of solving a system of nonlinear equations.
We write this problem mathematically as
r (x) 0,
(11.1)
CHAPTER 11.
NONLINEAR EQUATIONS
where r : IRn → IRn is a vector function, that is,
⎡
r1 (x)
⎢
⎢ r2 (x)
⎢
r (x) ⎢
..
⎢
.
⎣
rn (x)
⎤
⎥
⎥
⎥
⎥.
⎥
⎦
In this chapter, we assume that each function ri : IRn → IR, i 1, 2, . . . , n, is smooth. A
vector x ∗ for which (11.1) is satisfied is called a solution or root of the nonlinear equations.
A simple example is the system
r (x) x22 − 1
sin x1 − x2
0,
which is a system of n 2 equations with infinitely many solutions, two of which are
x ∗ (3π/2, −1)T and x ∗ (π/2, 1)T . In general, the system (11.1) may have no solutions,
a unique solution, or many solutions.
The techniques for solving nonlinear equations overlap in their motivation, analysis,
and implementation with optimization techniques discussed in earlier chapters. In both
optimization and nonlinear equations, Newton’s method lies at the heart of many important
algorithms. Features such as line searches, trust regions, and inexact solution of the linear
algebra subproblems at each iteration are important in both areas, as are other issues such
as derivative evaluation and global convergence.
Because some important algorithms for nonlinear equations proceed by minimizing
a sum of squares of the equations, that is,
min
x
n
ri2 (x),
i1
there are particularly close connections with the nonlinear least-squares problem discussed
in Chapter 10. The differences are that in nonlinear equations, the number of equations
equals the number of variables (instead of exceeding the number of variables, as is typically
the case in Chapter 10), and that we expect all equations to be satisfied at the solution, rather
than just minimizing the sum of squares. This point is important because the nonlinear
equations may represent physical or economic constraints such as conservation laws or
consistency principles, which must hold exactly in order for the solution to be meaningful.
Many applications require us to solve a sequence of closely related nonlinear systems,
as in the following example.
271
272
CHAPTER 11.
NONLINEAR EQUATIONS
❏ EXAMPLE 11.1
(RHEINBOLDT; SEE [212])
An interesting problem in control is to analyze the stability of an aircraft in response
to the commands of the pilot. The following is a simplified model based on force-balance
equations, in which gravity terms have been neglected.
The equilibrium equations for a particular aircraft are given by a system of 5 equations
in 8 unknowns of the form
F(x) ≡ Ax + φ(x) 0,
(11.2)
where F : IR8 → IR5 , the matrix A is given by
⎡
⎢
⎢
⎢
A⎢
⎢
⎢
⎣
−3.933
0.107
0.126
0
−9.99
0
−45.83
−7.64
0
−0.987
0
−22.95
0
−28.37
0
0
0.002
0
−0.235
0
5.67
0
−0.921
0
1.0
0
−1.0
0
−0.168
0
0
0
−1.0
0
−0.196
0
−0.0071
⎤
⎥
⎥
⎥
−6.51 ⎥
⎥,
⎥
0
⎦
0
and the nonlinear part is defined by
⎡
⎢
⎢
⎢
⎢
φ(x) ⎢
⎢
⎢
⎣
−0.727x2 x3 + 8.39x3 x4 − 684.4x4 x5 + 63.5x4 x2
0.949x1 x3 + 0.173x1 x5
−0.716x1 x2 − 1.578x1 x4 + 1.132x4 x2
−x1 x5
⎤
⎥
⎥
⎥
⎥
⎥.
⎥
⎥
⎦
x1 x4
The first three variables x1 , x2 , x3 , represent the rates of roll, pitch, and yaw, respectively, while x4 is the incremental angle of attack and x5 the sideslip angle. The last three
variables x6 , x7 , x8 are the controls; they represent the deflections of the elevator, aileron,
and rudder, respectively.
For a given choice of the control variables x6 , x7 , x8 we obtain a system of 5 equations
and 5 unknowns. If we wish to study the behavior of the aircraft as the controls are changed,
we need to solve a system of nonlinear equations with unknowns x1 , x2 , . . . , x5 for each
setting of the controls.
❐
Despite the many similarities between nonlinear equations and unconstrained and
least-squares optimization algorithms, there are also some important differences. To obtain quadratic convergence in optimization we require second derivatives of the objective
function, whereas knowledge of the first derivatives is sufficient in nonlinear equations.
CHAPTER 11.
NONLINEAR EQUATIONS
4
3
2
1
0
−1
−2
−3
−4
−3
−2
−1
0
1
2
3
Figure 11.1 The function r (x) sin(5x) − x has three roots.
Quasi-Newton methods are perhaps less useful in nonlinear equations than in optimization. In unconstrained optimization, the objective function is the natural choice of merit
function that gauges progress towards the solution, but in nonlinear equations various merit
functions can be used, all of which have some drawbacks. Line search and trust-region techniques play an equally important role in optimization, but one can argue that trust-region
algorithms have certain theoretical advantages in solving nonlinear equations.
Some of the difficulties that arise in trying to solve nonlinear equations can be
illustrated by a simple scalar example (n 1). Suppose we have
r (x) sin(5x) − x,
(11.3)
as plotted in Figure 11.1. From this figure we see that there are three solutions of the
problem r (x) 0, also known as roots of r , located at zero and approximately ±0.519148.
This situation of multiple solutions is similar to optimization problems where, for example,
a function may have more than one local minimum. It is not quite the same, however: In
the case of optimization, one of the local minima may have a lower function value than
the others (making it a “better” solution), while in nonlinear equations all solutions are
equally good from a mathematical viewpoint. (If the modeler decides that the solution
273
274
CHAPTER 11.
NONLINEAR EQUATIONS
found by the algorithm makes no sense on physical grounds, their model may need to be
reformulated.)
In this chapter we start by outlining algorithms related to Newton’s method and
examining their local convergence properties. Besides Newton’s method itself, these include Broyden’s quasi-Newton method, inexact Newton methods, and tensor methods.
We then address global convergence, which is the issue of trying to force convergence to
a solution from a remote starting point. Finally, we discuss a class of methods in which
an “easy” problem—one to which the solution is well known—is gradually transformed
into the problem F(x) 0. In these so-called continuation (or homotopy) methods, we
track the solution as the problem changes, with the aim of finishing up at a solution of
F(x) 0.
Throughout this chapter we make the assumption that the vector function r is continuously differentiable in the region D containing the values of x we are interested in. In
other words, the Jacobian J (x) (the matrix of first partial derivatives of r (x) defined in the
Appendix and in (10.3)) exists and is continuous. We say that x ∗ satisfying r (x ∗ ) 0 is a
degenerate solution if J (x ∗ ) is singular, and a nondegenerate solution otherwise.
11.1
LOCAL ALGORITHMS
NEWTON’S METHOD FOR NONLINEAR EQUATIONS
Recall from Theorem 2.1 that Newton’s method for minimizing f : IRn → IR forms a
quadratic model function by taking the first three terms of the Taylor series approximation
of f around the current iterate xk . The Newton step is the vector that minimizes this model.
In the case of nonlinear equations, Newton’s method is derived in a similar way, but with a
linear model, one that involves function values and first derivatives of the functions ri (x),
i 1, 2, . . . , m at the current iterate xk . We justify this strategy by referring to the following
multidimensional variant of Taylor’s theorem.
Theorem 11.1.
Suppose that r : IRn → IRn is continuously differentiable in some convex open set D and
that x and x + p are vectors in D. We then have that
r (x + p) r (x) +
1
J (x + t p) p dt.
(11.4)
0
We can define a linear model Mk ( p) of r (xk + p) by approximating the second term on the
right-hand-side of (11.4) by J (x) p, and writing
def
Mk ( p) r (xk ) + J (xk ) p.
(11.5)
11.1.
LOCAL ALGORITHMS
Newton’s method, in its pure form, chooses the step pk to be the vector for which Mk ( pk ) 0, that is, pk −J (xk )−1r (xk ). We define it formally as follows.
Algorithm 11.1 (Newton’s Method for Nonlinear Equations).
Choose x0 ;
for k 0, 1, 2, . . .
Calculate a solution pk to the Newton equations
J (xk ) pk −r (xk );
(11.6)
xk+1 ← xk + pk ;
end (for)
We use a linear model to derive the Newton step, rather than a quadratic model as in
unconstrained optimization, because the linear model normally has a solution and yields an
algorithm with rapid convergence properties. In fact, Newton’s method for unconstrained
optimization (see (2.15)) can be derived by applying Algorithm 11.1 to the nonlinear
equations ∇ f (x) 0. We see also in Chapter 18 that sequential quadratic programming
for equality-constrained optimization can be derived by applying Algorithm 11.1 to the
nonlinear equations formed by the first-order optimality conditions (18.3) for this problem.
Another connection is with the Gauss–Newton method for nonlinear least squares; the
formula (11.6) is equivalent to (10.23) in the usual case in which J (xk ) is nonsingular.
When the iterate xk is close to a nondegenerate root x ∗ , Newton’s method converges
superlinearly, as we show in Theorem 11.2 below. Potential shortcomings of the method
include the following.
• When the starting point is remote from a solution, Algorithm 11.1 can behave
erratically. When J (xk ) is singular, the Newton step may not even be defined.
• First-derivative information (the Jacobian matrix J ) may be difficult to obtain.
• It may be too expensive to find and calculate the Newton step pk exactly when n is
large.
• The root x ∗ in question may be degenerate, that is, J (x ∗ ) may be singular.
An example of a degenerate problem is the scalar function r (x) x 2 , which has a single
degenerate root at x ∗ 0. Algorithm 11.1, when started from any nonzero x0 , generates the
sequence of iterates
xk 1
x0 ,
2k
which converges to the solution 0, but only at a linear rate.
As we show later in this chapter, Newton’s method can be modified and enhanced in
various ways to get around most of these problems. The variants we describe form the basis
of much of the available software for solving nonlinear equations.
275
276
CHAPTER 11.
NONLINEAR EQUATIONS
We summarize the local convergence properties of Algorithm 11.1 in the following
theorem. For part of this result, we make use of a Lipschitz continuity assumption on the
Jacobian, by which we mean that there is a constant β L such that
J (x0 ) − J (x1 ) ≤ β L x0 − x1 ,
(11.7)
for all x0 and x1 in the domain in question.
Theorem 11.2.
Suppose that r is continuously differentiable in a convex open set D ⊂ IRn . Let x ∗ ∈ D
be a nondegenerate solution of r (x) 0, and let {xk } be the sequence of iterates generated by
Algorithm 11.1. Then when xk ∈ D is sufficiently close to x ∗ , we have
xk+1 − x ∗ o(xk − x ∗ ),
(11.8)
indicating local Q-superlinear convergence. When r is Lipschitz continuously differentiable
near x ∗ , we have for all xk sufficiently close to x ∗ that
xk+1 − x ∗ O(xk − x ∗ 2 ),
(11.9)
indicating local Q-quadratic convergence.
PROOF. Since r (x ∗ ) 0, we have from Theorem 11.1 that
r (xk ) r (xk ) − r (x ∗ ) J (xk )(xk − x ∗ ) + w(xk , x ∗ ),
(11.10)
where
w(xk , x ∗ ) 1
J (xk + t(x ∗ − xk )) − J (xk ) (xk − x ∗ ).
(11.11)
0
From (A.12) and continuity of J , we have
1
∗
∗
∗
w(xk , x ) [J (x + t(x − xk )) − J (xk )](xk − x ) dt 0
1
≤
J (x ∗ + t(x ∗ − xk )) − J (xk ) xk − x ∗ dt
∗
(11.12)
0
o(xk − x ∗ ).
Since J (x ∗ ) is nonsingular, there is a radius δ > 0 and a positive constant β ∗ such that for
all x in the ball B(x ∗ , δ) defined by
B(x ∗ , δ) {x | x − x ∗ ≤ δ},
(11.13)
11.1.
LOCAL ALGORITHMS
x ∈ D.
(11.14)
we have that
J (x)−1 ≤ β ∗
and
Assuming that xk ∈ B(x ∗ , δ), and recalling the definition (11.6), we multiply both sides of
(11.10) by J (xk )−1 to obtain
− pk (xk − x ∗ ) + J (xk )−1 o(xk − x ∗ ),
⇒ xk + pk − x ∗ o(xk − x ∗ ),
⇒ xk+1 − x ∗ o(xk − x ∗ ),
(11.15)
which yields (11.8).
When the Lipschitz continuity assumption (11.7) is satisfied, we can obtain a sharper
estimate for the remainder term w(xk , x ∗ ) defined in (11.11). By using (11.7) in (11.12), we
obtain
w(xk , x ∗ ) O(xk − x ∗ 2 ).
(11.16)
By multiplying (11.10) by J (xk )−1 as above, we obtain
− pk − (xk − x ∗ ) J (xk )−1 w(xk , x ∗ ),
so the estimate (11.9) follows as in (11.15).
INEXACT NEWTON METHODS
Instead of solving (11.6) exactly, inexact Newton methods use search directions pk
that satisfy the condition
rk + Jk pk ≤ ηk rk ,
for some ηk ∈ [0, η],
(11.17)
where η ∈ [0, 1) is a constant. As in Chapter 7, we refer to {ηk } as the forcing sequence.
Different methods make different choices of the forcing sequence, and they use different
algorithms for finding the approximate solutions pk . The general framework for this class
of methods can be stated as follows.
Framework 11.2 (Inexact Newton for Nonlinear Equations).
Given η ∈ [0, 1);
Choose x0 ;
for k 0, 1, 2, . . .
Choose forcing parameter ηk ∈ [0, η];
Find a vector pk that satisfies (11.17);
xk+1 ← xk + pk ;
end (for)
277
278
CHAPTER 11.
NONLINEAR EQUATIONS
The convergence theory for these methods depends only on the condition (11.17)
and not on the particular technique used to calculate pk . The most important methods
in this class, however, make use of iterative techniques for solving linear systems of the
form J p −r , such as GMRES (Saad and Schultz [273], Walker [302]) or other Krylovspace methods. Like the conjugate-gradient algorithm of Chapter 5 (which is not directly
applicable here, since the coefficient matrix J is not symmetric positive definite), these
methods typically require us to perform a matrix–vector multiplication of the form J d for
some d at each iteration, and to store a number of work vectors of length n. GMRES requires
an additional vector to be stored at each iteration, so must be restarted periodically (often
every 10 or 20 iterations) to keep memory requirements at a reasonable level.
The matrix–vector products J d can be computed without explicit knowledge of the
Jacobian J . A finite-difference approximation to J d that requires one evaluation of r (·)
is given by the formula (8.11). Calculation of J d exactly (at least, to within the limits of
finite-precision arithmetic) can be performed by using the forward mode of automatic
differentiation, at a cost of at most a small multiple of an evaluation of r (·). Details of this
procedure are given in Section 8.2.
We do not discuss the iterative methods for sparse linear systems here, but refer
the interested reader to Kelley [177] and Saad [272] for comprehensive descriptions and
implementations of the most interesting techniques. We prove a local convergence theorem
for the method, similar to Theorem 11.2.
Theorem 11.3.
Suppose that r is continuously differentiable in a convex open set D ⊂ IRn . Let x ∗ ∈ D
be a nondegenerate solution of r (x) 0, and let {xk } be the sequence of iterates generated by
the Framework 11.2. Then when xk ∈ D is sufficiently close to x ∗ , the following are true:
(i) If η in (11.17) is sufficiently small, the convergence of {xk } to x ∗ is Q-linear.
(ii) If ηk → 0, the convergence is Q-superlinear.
(iii) If, in addition, J (·) is Lipschitz continuous in a neighborhood of x ∗ and ηk O(rk ),
the convergence is Q-quadratic.
PROOF. We first rewrite (11.17) as
J (xk ) pk + r (xk ) vk ,
where vk ≤ ηk r (xk ).
(11.18)
Since x ∗ is a nondegenerate root, we have as in (11.14) that there is a radius δ > 0 such that
J (x)−1 ≤ β ∗ for some constant β ∗ and all x ∈ B(x ∗ , δ). By multiplying both sides of
(11.18) by J (xk )−1 and rearranging, we find that
pk + J (xk )−1r (xk ) J (xk )−1 vk ≤ β ∗ ηk r (xk ).
(11.19)
11.1.
LOCAL ALGORITHMS
r (x) J (x)(x − x ∗ ) + w(x, x ∗ ),
(11.20)
As in (11.10), we have that
def
where ρ(x) w(x, x ∗ )/x − x ∗ → 0 as x → x ∗ . By reducing δ if necessary, we have
from this expression that the following bound holds for all x ∈ B(x ∗ , δ):
r (x) ≤ 2J (x ∗ ) x − x ∗ + o(x − x ∗ ) ≤ 4J (x ∗ ) x − x ∗ .
(11.21)
We now set x xk in (11.20), and use (11.19) and (11.21) to obtain
xk + pk − x ∗ pk + J (xk )−1 (r (xk ) − w(xk , x ∗ ))
≤ β ∗ ηk r (xk ) + J (xk )−1 w(xk , x ∗ )
≤ 4J (x ∗ )β ∗ ηk + β ∗ ρ(xk ) xk − x ∗ .
(11.22)
By choosing xk close enough to x ∗ that ρ(xk ) ≤ 1/(4β ∗ ), and choosing η 1/(8J (x ∗ )β ∗ ), we have that the term in square brackets in (11.22) is at most 1/2. Hence,
since xk+1 xk + pk , this formula indicates Q-linear convergence of {xk } to x ∗ , proving
part (i).
Part (ii) follows immediately from the fact that the term in brackets in (11.22) goes to
zero as xk → x ∗ and ηk → 0. For part (iii), we combine the techniques above with the logic
of the second part of the proof of Theorem 11.2. Details are left as an exercise.
BROYDEN’S METHOD
Secant methods, also known as quasi-Newton methods, do not require calculation of
the Jacobian J (x). Instead, they construct their own approximation to this matrix, updating
it at each iteration so that it mimics the behavior of the true Jacobian J over the step just
taken. The approximate Jacobian, which we denote at iteration k by Bk , is then used to
construct a linear model analogous to (11.5), namely
Mk ( p) r (xk ) + Bk p.
(11.23)
We obtain the step by setting this model to zero. When Bk is nonsingular, we have the
following explicit formula (cf. (11.6)):
pk −Bk−1r (xk ).
(11.24)
The requirement that the approximate Jacobian should mimic the behavior of the
true Jacobian can be specified as follows. Let sk denote the step from xk to xk+1 , and let yk
279
280
CHAPTER 11.
NONLINEAR EQUATIONS
be the corresponding change in r , that is,
sk xk+1 − xk ,
yk r (xk+1 ) − r (xk ).
(11.25)
From Theorem 11.1, we have that sk and yk are related by the expression
yk 1
J (xk + tsk )sk dt ≈ J (xk+1 )sk + o(sk ).
(11.26)
0
We require the updated Jacobian approximation Bk+1 to satisfy the following equation,
which is known as the secant equation,
yk Bk+1 sk ,
(11.27)
which ensures that Bk+1 and J (xk+1 ) have similar behavior along the direction sk . (Note
the similarity with the secant equation (6.6) in quasi-Newton methods for unconstrained
optimization; the motivation is the same in both cases.) The secant equation does not say
anything about how Bk+1 should behave along directions orthogonal to sk . In fact, we can
view (11.27) as a system of n linear equations in n 2 unknowns, where the unknowns are
the components of Bk+1 , so for n > 1 the equation (11.27) does not determine all the
components of Bk+1 uniquely. (The scalar case of n 1 gives rise to the scalar secant
method; see (A.60).)
The most successful practical algorithm is Broyden’s method, for which the update
formula is
Bk+1 Bk +
(yk − Bk sk )skT
.
skT sk
(11.28)
The Broyden update makes the smallest possible change to the Jacobian (as measured by the
Euclidean norm Bk − Bk+1 2 ) that is consistent with (11.27), as we show in the following
Lemma.
Lemma 11.4 (Dennis and Schnabel [92, Lemma 8.1.1]).
Among all matrices B satisfying Bsk yk , the matrix Bk+1 defined by (11.28) minimizes
the difference B − Bk .
PROOF. Let B be any matrix that satisfies Bsk yk . By the properties of the Euclidean
norm (see (A.10)) and the fact that ss T /s T s 1 for any vector s (see Exercise 11.1), we
have
(yk − Bk sk )skT Bk+1 − Bk skT sk
T
(B − Bk )sk skT ≤ B − Bk sk sk B − Bk .
T
T
s s
s s k k
k k
11.1.
LOCAL ALGORITHMS
Hence, we have that
Bk+1 ∈ arg
min
B : yk Bsk
B − Bk ,
and the result is proved.
In the specification of the algorithm below, we allow a line search to be performed
along the search direction pk , so that sk αpk for some α > 0 in the formula (11.25). (See
below for details about line-search methods.)
Algorithm 11.3 (Broyden).
Choose x0 and a nonsingular initial Jacobian approximation B0 ;
for k 0, 1, 2, . . .
Calculate a solution pk to the linear equations
Bk pk −r (xk );
(11.29)
Choose αk by performing a line search along pk ;
xk+1 ← xk + αk pk ;
sk ← xk+1 − xk ;
yk ← r (xk+1 ) − r (xk );
Obtain Bk+1 from the formula (11.28);
end (for)
Under certain assumptions, Broyden’s method converges superlinearly, that is,
xk+1 − x ∗ o(xk − x ∗ ).
(11.30)
This local convergence rate is fast enough for most practical purposes, though not as fast as
the Q-quadratic convergence of Newton’s method.
We illustrate the difference between the convergence rates of Newton’s and Broyden’s
method with a small example. The function r : IR2 → IR2 defined by
r (x) (x1 + 3)(x23 − 7) + 18
sin(x2 e x1 − 1)
(11.31)
has a nondegenerate root at x ∗ (0, 1)T . We start both methods from the point
x0 (−0.5, 1.4)T , and use the exact Jacobian J (x0 ) at this point as the initial Jacobian
approximation B0 . Results are shown in Table 11.1.
Newton’s method clearly exhibits Q-quadratic convergence, which is characterized by
doubling of the exponent of the error at each iteration. Broyden’s method takes twice as
281
282
CHAPTER 11.
NONLINEAR EQUATIONS
Table 11.1 Convergence of Iterates in Broyden
and Newton Methods
Iteration k
0
1
2
3
4
5
6
7
8
xk − x ∗ 2
Broyden
Newton
0.64 × 100
0.64 × 100
−1
0.62 × 10
0.62 × 10−1
−3
0.52 × 10
0.21 × 10−3
−3
0.25 × 10
0.18 × 10−7
0.43 × 10−4
0.12 × 10−15
−6
0.14 × 10
0.57 × 10−9
0.18 × 10−11
0.87 × 10−15
Table 11.2 Convergence of Function Norms in
Broyden and Newton Methods
Iteration k
0
1
2
3
4
5
6
7
8
r (xk )2
Broyden
Newton
0.74 × 101
0.74 × 101
0
0.59 × 10
0.59 × 100
−2
0.20 × 10
0.23 × 10−2
−2
0.21 × 10
0.16 × 10−6
0.37 × 10−3
0.22 × 10−15
−5
0.12 × 10
0.49 × 10−8
0.15 × 10−10
0.11 × 10−18
many iterations as Newton’s, and reduces the error at a rate that accelerates slightly towards
the end. The function norms r (xk ) approach zero at a similar rate to the iteration errors
xk − x ∗ . As in (11.10), we have that
r (xk ) r (xk ) − r (x ∗ ) ≈ J (x ∗ )(xk − x ∗ ),
so by nonsingularity of J (x ∗ ), the norms of r (xk ) and (xk − x ∗ ) are bounded above and
below by multiples of each other. For our example problem (11.31), convergence of the
sequence of function norms in the two methods is shown in Table 11.2.
The convergence analysis of Broyden’s method is more complicated than that of
Newton’s method. We state the following result without proof.
Theorem 11.5.
Suppose the assumptions of Theorem 11.2 hold. Then there are positive constants and
δ such that if the starting point x0 and the starting approximate Jacobian B0 satisfy
x0 − x ∗ ≤ δ,
B0 − J (x ∗ ) ≤ ,
(11.32)
11.1.
LOCAL ALGORITHMS
the sequence {xk } generated by Broyden’s method (11.24), (11.28) is well-defined and converges
Q-superlinearly to x ∗ .
The second condition in (11.32)—that the initial Jacobian approximation B0 must
be close to the true Jacobian at the solution J (x ∗ )—is difficult to guarantee in practice.
In contrast to the case of unconstrained minimization, a good choice of B0 can be critical to the performance of the algorithm. Some implementations of Broyden’s method
recommend choosing B0 to be J (x0 ), or some finite-difference approximation to this
matrix.
The Broyden matrix Bk will be dense in general, even if the true Jacobian J is sparse.
Therefore, when n is large, an implementation of Broyden’s method that stores Bk as a full
n × n matrix may be inefficient. Instead, we can use limited-memory methods in which
Bk is stored implicitly in the form of a number of vectors of length n, while the system
(11.29) is solved by a technique based on application of the Sherman–Morrison–Woodbury
formula (A.28). These methods are similar to the ones described in Chapter 7 for large-scale
unconstrained optimization.
TENSOR METHODS
In tensor methods, the linear model Mk ( p) used by Newton’s method (11.5) is augmented with an extra term that aims to capture some of the nonlinear, higher-order,
behavior of r . By doing so, it achieves more rapid and reliable convergence to degenerate
roots, in particular, to roots x ∗ for which the Jacobian J (x ∗ ) has rank n − 1 or n − 2.
We give a broad outline of the method here, and refer to Schnabel and Frank [277] for
details.
We use M̂k ( p) to denote the model function on which tensor methods are based; this
function has the form
M̂k ( p) r (xk ) + J (xk ) p + 12 Tk pp,
(11.33)
where Tk is a tensor defined by n 3 elements (Tk )i jl whose action on a pair of arbitrary vectors
u and v in IRn is defined by
(Tk uv)i n
n (Tk )i jl u j vl .
j1 l1
If we followed the reasoning behind Newton’s method, we could consider building Tk from
the second derivatives of r at the point xk , that is,
(Tk )i jl [∇ 2ri (xk )] jl .
283
284
CHAPTER 11.
NONLINEAR EQUATIONS
For instance, in the example (11.31), we have that
(T (x)uv)1 u ∇ r1 (x)v u
T
2
T
0
3x22
3x22
6x2 (x1 + 3)
v
3x22 (u 1 v2 + u 2 v1 ) + 6x2 (x1 + 3)u 2 v2 .
However, use of the exact second derivatives is not practical in most instances. If we were to
store this information explicitly, about n 3 /2 memory locations would be needed, about n
times the requirements of Newton’s method. Moreover, there may be no vector p for which
M̂k ( p) 0, so the step may not even be defined.
Instead, the approach described in [277] defines Tk in a way that requires little
additional storage, but which gives M̂k some potentially appealing properties. Specifically,
Tk is chosen so that M̂k ( p) interpolates the function r (xk + p) at some previous iterates
visited by the algorithm. That is, we require that
M̂k (xk− j − xk ) r (xk− j ),
for j 1, 2, . . . , q,
(11.34)
for some integer q > 0. By substituting from (11.33), we see that Tk must satisfy the
condition
1
Ts s
2 k jk jk
r (xk− j ) − r (xk ) − J (xk )s jk ,
where
def
s jk xk− j − xk , j 1, 2, . . . , q.
In [277] it is shown that this condition can be ensured by choosing Tk so that its action on
arbitrary vectors u and v is
Tk uv q
a j (s Tjk u)(s Tjk v),
j1
where a j , j 1, 2, . . . , q, are vectors of length n. The number of interpolating points q
√
is typically chosen to be quite modest, usually less than n. This Tk can be stored in 2nq
locations, which contain the vectors a j and s jk for j 1, 2, . . . , q. Note the connection
between this idea and Broyden’s method, which also chooses information in the model
(albeit in the first-order part of the model) to interpolate the function value at the previous
iterate.
This technique can be refined in various ways. The points of interpolation can be
chosen to make the collection of directions s jk more linearly independent. There may still
not be a vector p for which M̂k ( p) 0, but we can instead take the step to be the vector that
11.2.
PRACTICAL METHODS
minimizes M̂k ( p)22 , which can be found by using a specialized least-squares technique.
There is no assurance that the step obtained in this way is a descent direction for the merit
function 12 r (x)2 (which is discussed in the next section), and in this case it can be replaced
by the standard Newton direction −Jk−1 rk .
11.2
PRACTICAL METHODS
We now consider practical variants of the Newton-like methods discussed above, in which
line-search and trust-region modifications to the steps are made in order to ensure better
global convergence behavior.
MERIT FUNCTIONS
As mentioned above, neither Newton’s method (11.6) nor Broyden’s method (11.24),
(11.28) with unit step lengths can be guaranteed to converge to a solution of r (x) 0 unless
they are started close to that solution. Sometimes, components of the unknown or function
vector or the Jacobian will blow up. Another, more exotic, kind of behavior is cycling, where
the iterates move between distinct regions of the parameter space without approaching a
root. An example is the scalar function
r (x) −x 5 + x 3 + 4x,
which has five nondegenerate roots. When started from the point x0 1, Newton’s method
produces a sequence of iterates that oscillates between 1 and −1 (see Exercise 11.3) without
converging to any of the roots.
The Newton and Broyden methods can be made more robust by using line-search and
trust-region techniques similar to those described in Chapters 3 and 4. Before describing
these techniques, we need to define a merit function, which is a scalar-valued function of x
that indicates whether a new iterate is better or worse than the current iterate, in the sense of
making progress toward a root of r . In unconstrained optimization, the objective function
f is itself a natural merit function; most algorithms for minimizing f require a decrease
in f at each iteration. In nonlinear equations, the merit function is obtained by combining
the n components of the vector r in some way.
The most widely used merit function is the sum of squares, defined by
f (x) 12 r (x)2 1
2
n
ri2 (x).
(11.35)
i1
(The factor 1/2 is introduced for convenience.) Any root x ∗ of r obviously has f (x ∗ ) 0,
and since f (x) ≥ 0 for all x, each root is a minimizer of f . However, local minimizers of
f are not roots of r if f is strictly positive at the point in question. Still, the merit function
285
286
CHAPTER 11.
NONLINEAR EQUATIONS
6
5
4
3
2
1
0
−3
Figure 11.2
−2
−1
0
1
2
3
Plot of 12 [sin(5x) − x]2 , showing its many local minima.
(11.35) has been used successfully in many applications and is implemented in a number of
software packages.
The merit function for the example (11.3) is plotted in Figure 11.2. It shows three
local minima corresponding to the three roots, but there are many other local minima (for
example, those at around ±1.53053). Local minima like these that are not roots of f satisfy
an interesting property. Since
∇ f (x ∗ ) J (x ∗ )T r (x ∗ ) 0,
(11.36)
we can have r (x ∗ ) 0 only if J (x ∗ ) is singular.
Since local minima for the sum-of-squares merit function may be points of attraction
for the algorithms described in this section, global convergence results for the algorithms
discussed here are less satisfactory than for similar algorithms applied to unconstrained
optimization.
Other merit functions are also used in practice. One such is the 1 norm merit function
defined by
f 1 (x) r (x)1 m
|ri (x)|.
i1
This function is studied in Chapters 17 and 18 in the context of algorithms for constrained
optimization.
11.2.
PRACTICAL METHODS
LINE SEARCH METHODS
We can obtain algorithms with global convergence properties by applying the linesearch approach of Chapter 3 to the sum-of-squares merit function f (x) 12 r (x)2 .
When it is well defined, the Newton step
J (xk ) pk −r (xk )
(11.37)
is a descent direction for f (·) whenever rk 0, since
pkT ∇ f (xk ) − pkT JkT rk −rk 2 < 0.
(11.38)
Step lengths αk are chosen by one of the procedures of Chapter 3, and the iterates are defined
by the formula
xk+1 xk + αk pk ,
k 0, 1, 2, . . . .
(11.39)
For the case of line searches that choose αk to satisfy the Wolfe conditions (3.6), we have the
following convergence result, which follows directly from Theorem 3.2.
Theorem 11.6.
Suppose that J (x) is Lipschitz continuous in a neighborhood D of the level set L {x : f (x) ≤ f (x0 )}, and that J (x) and r (x) are bounded above on D. Suppose that a
line-search algorithm (11.39) is applied to f , where the search directions pk satisfy pkT ∇ f k < 0
while the step lengths αk satisfy the Wolfe conditions (3.6). Then we have that the Zoutendijk
condition holds, that is,
cos2 θk JkT rk 2 < ∞,
k≥0
where
cos θk − pkT ∇ f (xk )
.
pk ∇ f (xk )
(11.40)
We omit the proof, which verifies that ∇ f is Lipschitz continuous on D and that f is
bounded below (by 0) on D, and then applies Theorem 3.2.
Provided that the sequence of iterates satisfies
cos θk ≥ δ,
for some δ ∈ (0, 1) and all k sufficiently large,
(11.41)
Theorem 11.6 guarantees that JkT rk → 0, meaning that the iterates approach stationarity of
the merit function f . Moreover, if we know that J (xk )−1 is bounded then we must have
rk → 0.
287
288
CHAPTER 11.
NONLINEAR EQUATIONS
We now investigate the values of cos θk for the directions generated by the Newton
and inexact Newton methods. From (11.40) and (11.38), we have for the exact Newton step
(11.6) that
cos θk −
pkT ∇ f (xk )
rk 2
1
1
.
≥
−1
pk ∇ f (xk )
κ(Jk )
Jk rk JkT rk JkT Jk−1 (11.42)
When pk is an inexact Newton direction—that is, one that satisfies the condition
(11.17)—we have that
rk + Jk pk 2 ≤ ηk2 rk 2 ⇒ 2 pkT JkT rk + rk 2 + Jk pk 2 ≤ η2 rk 2
⇒ pkT ∇ f k pkT JkT rk ≤ [(η2 − 1)/2]rk 2 .
Meanwhile,
pk ≤ Jk−1 [rk + Jk pk + rk ] ≤ Jk−1 (η + 1)rk ,
and
∇ f k JkT rk ≤ Jk rk .
By combining these estimates, we obtain
cos θk −
pkT ∇ f k
1 − η2
1−η
≥
.
≥
−1
pk ∇ f k 2κ(J
2Jk Jk (1 + η)
k)
We conclude that a bound of the form (11.41) is satisfied both for the exact and inexact
Newton methods, provided that the condition number κ(Jk ) is bounded.
When κ(Jk ) is large, however, this lower bound is close to zero, and use of the Newton
direction may cause poor performance of the algorithm. In fact, the following example
shows that condition cos θk can converge to zero, causing the algorithm to fail. This example
highlights a fundamental weakness of the line-search approach.
❏ EXAMPLE 11.2
(POWELL [241])
Consider the problem of finding a solution of the nonlinear system
⎡
r (x) ⎣
x1
⎤
⎦,
10x1
+ 2x22
(x1 + 0.1)
(11.43)
11.2.
PRACTICAL METHODS
with unique solution x ∗ 0. We try to solve this problem using the Newton iteration (11.37),
(11.39) where αk is chosen to minimize f along pk . It is proved in [241] that, starting from
the point (3, 1)T , the iterates converge to (1.8016, 0)T (to four digits of accuracy). However,
this point is not a solution of (11.43). In fact, it is not even a stationary point of f , and a
step from this point in the direction −∇ f will produce a decrease in both components of
r . To verify these claims, note that the Jacobian of r , which is
⎡
1
J (x) ⎣
1
(x1 + 0.1)2
0
4x2
⎤
⎦,
is singular at all x for which x2 0. For such points, we have
⎡
∇ f (x) ⎣
x1 +
⎤
10x1
(x1 + 0.1)3 ⎦ ,
0
so that the gradient points in the direction of the positive x1 axis whenever x1 > 0. The
point (1.8016, 0)T is therefore not a stationary point of f .
For this example, a calculation shows that the Newton step generated from an iterate
that is close to (but not quite on) the x1 axis tends to be parallel to the x2 axis, making it
nearly orthogonal to the gradient ∇ f (x). That is, cos θk for the Newton direction may be
arbitrarily close to zero.
❐
In this example, a Newton method with exact line searches is attracted to a point of
no interest at which the Jacobian is singular. Since systems of nonlinear equations often
contain singular points, this behavior gives cause for concern.
To prevent this undesirable behavior and ensure that (11.41) holds, we may have to
modify the Newton direction. One possibility is to add some multiple λk I of the identity to
JkT Jk , and define the step pk to be
pk −(JkT Jk + λk I )−1 JkT rk .
(11.44)
For any λk > 0 the matrix in parentheses is nonsingular, and if λk is bounded away from zero,
a condition of the form (11.41) is satisfied. Therefore, some practical algorithms choose λk
adaptively to ensure that the matrix in (11.44) does not approach singularity. This approach
is analogous to the classical Levenberg-Marquardt algorithm discussed in Chapter 10. To
implement it without forming JkT Jk explicitly and performing trial Cholesky factorizations
of the matrices (JkT Jk + λI ), we can use the technique (10.36) illustrated earlier for the
least-squares case. This technique uses the fact that the Cholesky factor of (JkT Jk + λI ) is
289
290
CHAPTER 11.
NONLINEAR EQUATIONS
identical to R T , where R is the upper triangular factor from the QR factorization of the
matrix
Jk
√
.
(11.45)
λI
A combination of Householder and Givens transformations can be used, as for (10.36), and
the savings noted in the discussion following (10.36) continue to hold if we need to perform
this calculation for several candidate values of λk .
The drawback of this Levenberg-Marquardt approach is that it is difficult to choose
λk . If too large, we can destroy the fast rate of convergence of Newton’s method. (Note
that pk approaches a multiple of −JkT rk as λk ↑ ∞, so the step becomes small and tends
to point in the steepest-descent direction for f .) If λk is too small, the algorithm can be
inefficient in the presence of Jacobian singularities. A more satisfactory approach is to follow
the trust-region approach described below, which chooses λk indirectly.
We conclude by specifying an algorithm based on Newton-like steps and line searches
that regularizes the step calculations where necessary. Several details are deliberately left
vague; we refer the reader to the papers cited above for details.
Algorithm 11.4 (Line Search Newton-like Method).
Given c1 , c2 with 0 < c1 < c2 < 12 ;
Choose x0 ;
for k 0, 1, 2, . . .
Calculate a Newton-like step from (11.6) (regularizing with (11.44)
if Jk appears to be near-singular), or (11.17) or (11.24);
if α 1 satisfies the Wolfe conditions (3.6)
Set αk 1;
else
Perform a line search to find αk > 0 that satisfies (3.6);
end (if)
xk+1 ← xk + αk pk ;
end (for)
TRUST-REGION METHODS
The most widely used trust-region methods for nonlinear equations simply apply Algorithm 4.1 from Chapter 4 to the merit function f (x) 12 r (x)22 , using
Bk J (xk )T J (xk ) as the approximate Hessian in the model function m k ( p), which is
defined as follows:
m k ( p) 12 rk + Jk p22 f k + p T JkT rk + 12 p T JkT Jk pk .
11.2.
PRACTICAL METHODS
The step pk is generated by finding an approximate solution of the subproblem
min m k ( p),
p
subject to p ≤ k ,
(11.46)
where k is the radius of the trust region. The ratio ρk of actual to predicted reduction (see
(4.4)), which plays a critical role in many trust-region algorithms, is therefore
ρk r (xk )2 − r (xk + pk )2
.
r (xk )2 − r (xk ) + J (xk ) pk 2
(11.47)
We can state the trust-region framework that results from this model as follows.
Algorithm 11.5 (Trust-Region Method for Nonlinear Equations).
¯ > 0, 0 ∈ (0, ),
¯ and η ∈ 0, 1 :
Given 4
for k 0, 1, 2, . . .
Calculate pk as an (approximate) solution of (11.46);
Evaluate ρk from (11.47);
if ρk < 14
k+1 14 pk ;
else
if ρk > 34 and pk k
¯
k+1 min(2k , );
else
k+1 k ;
end (if)
end (if)
if ρk > η
xk+1 xk + pk ;
else
xk+1 xk ;
end (if)
end (for).
The dogleg method is a special case of the trust-region algorithm, Algorithm 4.1,
that constructs an approximate solution to (11.46) based on the Cauchy point pkC and the
unconstrained minimizer of m k . The Cauchy point is
pkC −τk (k /JkT rk )JkT rk ,
(11.48)
2
3
τk min 1, JkT rk 3 /(k rkT Jk (JkT Jk )JkT rk ) ;
(11.49)
where
291
292
CHAPTER 11.
NONLINEAR EQUATIONS
By comparing with the general definition (4.11), (4.12) we see that it is not necessary to
consider the case of an indefinite Hessian approximation in m k ( p), since the model Hessian
JkT Jk that we use is positive semidefinite. The unconstrained minimizer of m k ( p) is unique
when Jk is nonsingular. In this case, we denote it by pkJ and write
pkJ −(JkT Jk )−1 (JkT rk ) −Jk−1 rk .
The selection of pk in the dogleg method proceeds as follows.
Procedure 11.6 (Dogleg).
Calculate pkC ;
if pkC k
pk ← pkC ;
else
Calculate pkJ ;
pk ← pkC + τ ( pkJ − pkC ), where τ is the largest value in [0, 1]
such that pk ≤ k ;
end (if).
Lemma 4.2 shows that when Jk is nonsingular, the vector pk chosen above is the
minimizer of m k along the piecewise linear path that leads from the origin to the Cauchy
point and then to the unconstrained minimizer pkJ . Hence, the reduction in model function
at least matches the reduction obtained by the Cauchy point, which can be estimated by
specializing the bound (4.20) to the least-squares case by writing
J T rk ,
m k (0) − m k ( pk ) ≥ c1 JkT rk min k , kT
Jk Jk (11.50)
where c1 is some positive constant.
From Theorem 4.1, we know that the exact solution of (11.46) has the form
pk −(JkT Jk + λk I )−1 JkT rk ,
(11.51)
for some λk ≥ 0, and that λk 0 if the unconstrained solution pkJ satisfies pkJ ≤ k . (Note
that (11.51) is identical to the formula (10.34a) from Chapter 10. In fact, the Levenberg–
Marquardt approach for nonlinear equations is a special case of the same algorithm for
nonlinear least-squares problems.) The Levenberg–Marquardt algorithm uses the techniques
of Section 4.3 to search for the value of λk that satisfies (11.51). The procedure described
in the “exact” trust-region algorithm, Algorithm 4.3, is based on Cholesky factorizations,
but as in Chapter 10, we can replace these by specialized algorithms to compute the QR
factorization of the matrix (11.45). Even if the exact λk corresponding to the solution of
(11.46) is not found, the pk calculated from (11.51) will still yield global convergence if it
11.2.
PRACTICAL METHODS
satisfies the condition (11.50) for some value of c1 , together with
pk ≤ γ k ,
for some constant γ ≥ 1.
(11.52)
The dogleg method requires just one linear system to be solved per iteration, whereas
methods that search for the exact solution of (11.46) require several such systems to be
solved. As in Chapter 4, there is a tradeoff to be made between the amount of effort to spend
on each iteration and the total number of function and derivative evaluations required.
We can also consider alternative trust-region approaches that are based on different
merit functions and different definitions of the trust region. An algorithm based on the 1
merit function with an ∞ -norm trust region gives rise to subproblems of the form
min Jk p + rk 1
p
subject to p∞ ≤ ,
(11.53)
which can be formulated and solved using linear programming techniques. This approach
is closely related to the S1 QP and SLQP approaches for nonlinear programming discussed
in Section 18.5.
Global convergence results of Algorithm 11.5 when the steps pk satisfy (11.50) and
(11.52) are given in the following theorem, which can be proved by referring directly to
Theorems 4.5 and 4.6. The first result is for η 0, in which the algorithm accepts all steps
that produce a decrease in the merit function f k , while the second (stronger) result requires
a strictly positive choice of η.
Theorem 11.7.
Suppose that J (x) is Lipschitz continuous and that J (x) is bounded above in a
neighborhood D of the level set L {x : f (x) ≤ f (x0 )}. Suppose in addition that all
approximate solutions of (11.46) satisfy the bounds (11.50) and (11.52). Then if η 0 in
Algorithm 11.5, we have that
lim inf JkT rk 0,
k→∞
while if η ∈ 0, 14 , we have
lim JkT rk 0.
k→∞
We turn now to local convergence of the trust-region algorithm for the case in which
the subproblem (11.46) is solved exactly. We assume that the sequence {xk } converges to
a nondegenerate solution x ∗ of the nonlinear equations r (x) 0. The significance of
this result is that the algorithmic enhancements needed for global convergence do not, in
well-designed algorithms, interfere with the fast local convergence properties described in
Section 11.1.
293
294
CHAPTER 11.
NONLINEAR EQUATIONS
Theorem 11.8.
Suppose that the sequence {xk } generated by Algorithm 11.5 converges to a nondegenerate
solution x ∗ of the problem r (x) 0. Suppose also that J (x) is Lipschitz continuous in an open
neighborhood D of x ∗ and that the trust-region subproblem (11.46) is solved exactly for all
sufficiently large k. Then the sequence {xk } converges quadratically to x ∗ .
PROOF. We prove this result by showing that there is an index K such that the trust-region
radius is not reduced further after iteration K ; that is, k ≥ K for all k ≥ K . We then
show that the algorithm eventually takes the pure Newton step at every iteration, so that
quadratic convergence follows from Theorem 11.2.
Let pk denote the exact solution of (11.46). Note first that pk will simply be the unconstrained Newton step −Jk−1 rk whenever this step satisfies the trust-region bound. Otherwise,
we have Jk−1 rk > k , while the solution pk satisfies pk ≤ k . In either case, we have
pk ≤ Jk−1 rk .
(11.54)
We consider the ratio ρk of actual to predicted reduction defined by (11.47). We have
directly from the definition that
rk + Jk pk 2 − r (xk + pk )2 .
|1 − ρk | ≤
r (xk )2 − r (xk ) + J (xk ) pk 2
(11.55)
From Theorem 11.1, we have for the second term in the numerator that
r (xk + pk )2 [r (xk ) + J (xk ) pk ] + w(xk , xk + pk )2 ,
(11.56)
where w(·, ·) is defined as in (11.11). Because of Lipschitz continuity of J with Lipschitz
constant β L (11.7), we have
1
w(xk , xk + pk ) ≤
J (xk + t pk ) − J (xk ) pk dt
0
1
≤
β L pk 2 dt (β L /2) pk 2 ,
0
so that using (11.56) and the fact that rk + Jk pk ≤ rk f (xk )1/2 (since pk is the
solution of (11.46)), we can bound the numerator as follows:
rk + Jk pk 2 − r (xk + pk )2 ≤ 2rk + Jk pk w(xk , xk + pk ) + w(xk , xk + pk )2
≤ f (xk )1/2 β L pk 2 + (β L /2)2 pk 4
≤ (xk ) pk 2 ,
(11.57)
11.2.
PRACTICAL METHODS
where we define
(xk ) f (xk )1/2 β L + (β L /2)2 pk 2 .
Since xk → x ∗ by assumption, it follows that f (xk ) → 0 and rk → 0. Because x ∗ is a
nondegenerate root, we have as in (11.14) that J (xk )−1 ≤ β ∗ for all k sufficiently large,
so from (11.54), we have
pk ≤ Jk−1 rk ≤ β ∗ rk → 0.
(11.58)
Hence, (xk ) → 0.
Turning now to the denominator of (11.55), we define p̄k to be a step of the same
length as the solution pk in the Newton direction −Jk−1rk , that is,
p̄k −
pk −1
J rk .
Jk−1rk k
Since p̄k is feasible for (11.46), and since pk is optimal for this subproblem, we have
rk − rk + Jk pk ≥ rk − rk −
2
2
2
2
pk rk −1
Jk rk pk pk 2
rk 2 − −1 2 rk 2
−1
Jk rk Jk rk pk ≥
rk 2 ,
Jk−1 rk 2
where for the last inequality we have used (11.54). By using (11.58) again, we have from this
bound that
rk 2 − rk + Jk pk 2 ≥
pk 1
rk 2 ≥ ∗ pk rk .
β
Jk−1rk (11.59)
By substituting (11.57) and (11.59) into (11.55), and then applying (11.58) again, we have
|1 − ρk | ≤
β ∗ (xk ) pk 2
≤ (β ∗ )2 (xk ) → 0.
pk rk (11.60)
Therefore, for all k sufficiently large, we have ρk > 14 , and so the trust region radius k will
not be increased beyond this point. As claimed, there is an index K such that
k ≥ K ,
for all k ≥ K .
295
296
CHAPTER 11.
NONLINEAR EQUATIONS
Since Jk−1rk ≤ β ∗ rk → 0, the Newton step −Jk−1 rk will eventually be smaller
than K (and hence k ), so it will eventually always be accepted as the solution of (11.46).
The result now follows from Theorem 11.2.
We can replace the assumption that xk → x ∗ with an assumption that the nondegenerate solution x ∗ is just one of the limit points of the sequence. (In fact, this condition
implies that xk → x ∗ ; see Exercise 11.9.)
11.3
CONTINUATION/HOMOTOPY METHODS
MOTIVATION
We mentioned above that Newton-based methods all suffer from one shortcoming:
Unless J (x) is nonsingular in the region of interest—a condition that often cannot be
guaranteed—they are in danger of converging to a local minimum of the merit function
rather that is not a solution of the nonlinear system. Continuation methods, which we
outline in this section, are more likely to converge to a solution of r (x) 0 in difficult cases.
Their underlying motivation is simple to describe: Rather than dealing with the original
problem r (x) 0 directly, we set up an “easy” system of equations for which the solution
is obvious. We then gradually transform the easy system into the original system r (x), and
follow the solution as it moves from the solution of the easy problem to the solution of the
original problem.
One simple way to define the so-called homotopy map H (x, λ) is as follows:
H (x, λ) λr (x) + (1 − λ)(x − a),
(11.61)
where λ is a scalar parameter and a ∈ IRn is a fixed vector. When λ 0, (11.61) defines the
artificial, easy problem H (x, 0) x − a, whose solution is obviously x a. When λ 1,
we have H (x, 1) r (x), the original system of equations.
To solve r (x) 0, consider the following algorithm: First, set λ 0 in (11.61) and set
x a. Then, increase λ from 0 to 1 in small increments, and for each value of λ, calculate
the solution of the system H (x, λ) 0. The final value of x corresponding to λ 1 will
solve the original problem r (x) 0.
This naive approach sounds plausible, and Figure 11.3 illustrates a situation in which
it would be successful. In this figure, there is a unique solution x of the system H (x, λ) 0
for each value of λ in the range [0, 1]. The trajectory of points (x, λ) for which H (x, λ) 0
is called the zero path.
Unfortunately, however, the approach often fails, as illustrated in Figure 11.4. Here,
the algorithm follows the lower branch of the curve from λ 0 to λ λT , but it then loses
the trail unless it is lucky enough to jump to the top branch of the path. The value λT is
11.3.
CONTINUATION/HOMOTOPY METHODS
x
λ
0
Figure 11.3
1
Plot of a zero path: Trajectory of points (x, λ) with H (x, λ) 0.
. .
(x, λ )
x
0
λ
λT
1
Figure 11.4 Zero path with turning points. The path joining (a, 0) to (x ∗ , 1) cannot
be followed by increasing λ monotonically from 0 to 1.
known as a turning point, since at this point we can follow the path smoothly only if we no
longer insist on increasing λ at every step. In fact, practical continuation methods work by
doing exactly as Figure 11.4 suggests, that is, they follow the zero path explicitly, even if this
means allowing λ to decrease from time to time.
PRACTICAL CONTINUATION METHODS
In one practical technique, we model the zero path by allowing both x and λ to be
functions of an independent variable s that represents arc length along the path. That is,
297
298
CHAPTER 11.
NONLINEAR EQUATIONS
(x(s), λ(s)) is the point that we arrive at by traveling a distance s along the path from the
initial point (x(0), λ(0)) (a, 0). Because we have that
H (x(s), λ(s)) 0,
for all s ≥ 0,
we can take the total derivative of this expression with respect to s to obtain
∂
∂
H (x, λ)ẋ +
H (x, λ)λ̇ 0,
∂x
∂λ
where (ẋ, λ̇) d x dλ
.
,
ds ds
(11.62)
The vector (ẋ(s), λ̇(s)) is the tangent vector to the zero path, as we illustrate in Figure 11.4.
From (11.62), we see that it lies in the null space of the n × (n + 1) matrix
∂
H (x, λ)
∂x
∂
H (x, λ)
∂λ
.
(11.63)
When this matrix has full rank, its null space has dimension 1, so to complete the definition
of (ẋ, λ̇) in this case, we need to assign it a length and direction. The length is fixed by
imposing the normalization condition
ẋ(s)2 + |λ̇(s)|2 1,
for all s,
(11.64)
which ensures that s is the true arc length along the path from (0, a) to (x(s), λ(s)). We need
to choose the sign to ensure that we keep moving forward along the zero path. A heuristic
that works well is to choose the sign so that the tangent vector (ẋ, λ̇) at the current value of
s makes an angle of less than π/2 with the tangent point at the previous value of s.
We can outline the complete procedure for computing (ẋ, λ̇) as follows:
Procedure 11.7 (Tangent Vector Calculation).
Compute a vector in the null space of (11.63) by performing a QR
factorization with column pivoting,
QT
∂
H (x, λ)
∂x
∂
H (x, λ)
∂λ
R
w
where Q is n × n orthogonal, R is n × n upper triangular, is
an (n + 1) × (n + 1) permutation matrix, and w ∈ IRn .
Set
v
R −1 w
−1
;
,
11.3.
CONTINUATION/HOMOTOPY METHODS
Set (ẋ, λ̇) ±v/v2 , where the sign is chosen to satisfy the angle
criterion mentioned above.
Details of the QR factorization procedure are given in the Appendix.
Since we can obtain the tangent at any given point (x, λ) and since we know the initial
point (x(0), λ(0)) (a, 0), we can trace the zero path by calling a standard initial-value
first-order ordinary differential equation solver, terminating the algorithm when it finds a
value of s for which λ(s) 1.
A second approach for following the zero path is quite similar to the one just described,
except that it takes an algebraic viewpoint instead of a differential-equations viewpoint.
Given a current point (x, λ), we compute the tangent vector (ẋ, λ̇) as above, and take a
small step (of length , say) along this direction to produce a “predictor” point (x P , λ P );
that is,
(x P , λ P ) (x, λ) + (ẋ, λ̇).
Usually, this new point will not lie exactly on the zero path, so we apply some “corrector”
iterations to bring it back to the path, thereby identifying a new iterate (x + , λ+ ) that satisfies
H (x + , λ+ ) 0. (This process is illustrated in Figure 11.5.) During the corrections, we
choose a component of the predictor step (x P , λ P )—one of the components that has been
changing most rapidly during the past few steps—and hold this component fixed during
the correction process. If the index of this component is i, and if we use a pure Newton
corrector process (often adequate, since (x P , λ P ) is usually quite close to the target point
x
(x+, λ+)
(x, λ )
(x P, λP)
λ
Figure 11.5 The algebraic predictor–corrector procedure, using λ as the fixed
variable in the correction process.
299
300
CHAPTER 11.
NONLINEAR EQUATIONS
(x + , λ+ )), the steps will have the form
⎡
∂H
⎣ ∂x
⎤
∂H
δx
−H
,
∂λ ⎦
δλ
0
ei
where the quantities ∂ H/∂ x, ∂ H/∂λ, and H are evaluated at the latest point of the corrector
process. The last row of this system serves to fix the ith component of (δx, δλ) at zero; the
vector ei ∈ IRn+1 is a vector with n + 1 components containing all zeros, except for a 1
in the location i that corresponds to the fixed component. Note that in Figure 11.5 the λ
component is chosen to be fixed on the current iteration. On the following iteration, it may
be more appropriate to choose x as the fixed component, as we reach the turning point in λ.
The two variants on path-following described above are able to follow curves like
those depicted in Figure 11.4 to a solution of the nonlinear system. They rely, however, on
the n × (n + 1) matrix in (11.63) having full rank for all (x, λ) along the path, so that the
tangent vector is well-defined. The following result shows that full rank is guaranteed under
certain assumptions.
Theorem 11.9 (Watson [305]).
Suppose that r is twice continuously differentiable. Then for almost all vectors a ∈ IRn ,
there is a zero path emanating from (0, a) along which the n × (n + 1) matrix (11.63) has full
rank. If this path is bounded for λ ∈ [0, 1), then it has an accumulation point (x̄, 1) such that
r (x̄) 0. Furthermore, if the Jacobian J (x̄) is nonsingular, the zero path between (a, 0) and
(x̄, 1) has finite arc length.
The theorem assures us that unless we are unfortunate in the choice of a, the algorithms
described above can be applied to obtain a path that either diverges or else leads to a point
x̄ that is a solution of the original nonlinear system if J (x̄) is nonsingular. More detailed
convergence results can be found in Watson [305] and the references therein.
We conclude with an example to show that divergence of the zero path—the less
desirable outcome of Theorem 11.9—can happen even for innocent-looking problems.
❏ EXAMPLE 11.3
Consider the system r (x) x 2 − 1, for which there are two nondegenerate solutions
+1 and −1. Suppose we choose a −2 and attempt to apply a continuation method to the
function
H (x, λ) λ(x 2 − 1) + (1 − λ)(x + 2) λx 2 + (1 − λ)x + (2 − 3λ),
(11.65)
obtained by substituting into (11.61). The zero paths for this function are plotted in
Figure 11.6. As can be seen from that diagram, there is no zero path that joins (−2, 0)
11.3.
CONTINUATION/HOMOTOPY METHODS
2
0
−2
x
−4
−6
−8
−10
−12
0
0.2
0.4
0.6
lambda
0.8
1
Figure 11.6 Zero paths for the example in which H (x, λ) λ(x 2 −1)+(1−λ)(x+2).
There is no continuous zero path from λ 0 to λ 1.
to either (1, 1) or (−1, 1), so the continuation methods fail on this example. We can find the
values of λ for which no solution exists by using the formula for a quadratic root to obtain
x
−(1 − λ) ±
(1 − λ)2 − 4λ(2 − 3λ)
.
2λ
Now, when the term in the square root is negative, the corresponding values of x are complex,
that is, there are no real roots x. It is easy to verify that such is the case when
λ∈
√
√ 5−2 3 5+2 3
,
≈ (0.118, 0.651).
13
13
Note that the zero path starting from (−2, 0) becomes unbounded, which is one of the
possible outcomes of Theorem 11.9.
❐
This example indicates that continuation methods may fail to produce a solution even
to a fairly simple system of nonlinear equations. However, it is generally true that they are
more reliable than the merit-function methods described earlier in the chapter. The extra
robustness comes at a price, since continuation methods typically require significantly more
computational effort than the merit-function methods.
301
302
CHAPTER 11.
NONLINEAR EQUATIONS
NOTES AND REFERENCES
Nonlinear differential equations and integral equations are a rich source of nonlinear
equations. When formulated as finite-dimensional nonlinear equations, the unknown vector
x is a discrete approximation to the (infinite-dimensional) solution. In other applications,
the vector x is intrinsically finite-dimensional; it may represent the quantities of materials
to be transported between pairs of cities in a distribution network, for instance. In all cases,
the equations ri enforce consistency, conservation, and optimality principles in the model.
Moré [212] and Averick et al. [10] discuss a number of interesting practical applications.
For analysis of the convergence of Broyden’s method, including proofs of Theorem 11.5, see Dennis and Schnabel [92, Chapter 8] and Kelley [177, Chapter 6]. Details on a
limited-memory implementation of Broyden’s method are given by Kelley [177, Section 7.3].
Example 11.2 and the algorithm described by Powell [241] have been influential
beyond the field of nonlinear equations. The example shows that a line-search method
may not be able to achieve sufficient decrease, whereas the Cauchy step in the trust-region
approach is designed to guarantee that this condition holds and hence that reasonable
convergence properties are guaranteed. The dogleg algorithm proposed in [241] can be
viewed as one of the first modern trust-region methods.
✐
EXERCISES
✐
11.1 Show that for any vector s ∈ IRn , we have
T
ss s T s 1,
where · denotes the Euclidean matrix norm.
✐
11.2 Consider the function r : IR → IR defined by r (x) x q , where q is an integer
greater than 2. Note that x ∗ 0 is the sole root of this function and that it is degenerate.
Show that Newton’s method converges Q-linearly, and find the value of the convergence
ratio r in (A.34).
✐
11.3 Show that Newton’s method applied to the function r (x) −x 5 + x 3 + 4x
starting from x0 1 produces the cyclic behavior described in the text. Find the roots of
this function, and check that they are nondegenerate.
✐
11.4 For the scalar function r (x) sin(5x) − x, show that the sum-of-squares merit
function has infinitely many local minima, and find a general formula for such points.
✐
11.5 When r : IRn → IRn , show that the function
φ(λ) (J T J + λI )−1 J T r 11.3.
CONTINUATION/HOMOTOPY METHODS
is monotonically decreasing in λ unless J T r 0. (Hint: Use the singular-value
decomposition of J .)
✐
11.6 Prove part (iii) of Theorem 11.3.
✐
11.7 Consider a line-search Newton method in which the step length αk is chosen to
be the exact minimizer of the merit function f (·); that is,
αk arg min f (xk − α Jk−1rk ).
α
Show that if J (x) is nonsingular at the solution x ∗ , then αk → 1 as xk → x ∗ .
✐
11.8 Let J ∈ IRn×m and r ∈ IRn and suppose that J J T r 0. Show that J T r 0.
(Hint: This doesn’t even take one line!)
✐
11.9 Suppose we replace the assumption of xk → x ∗ in Theorem 11.8 by an assumption that the nondegenerate solution x ∗ is a limit point of x ∗ . By adding some logic to the
proof of this result, show that in fact x ∗ is the only possible limit point of the sequence.
−1
rk+1 ≤ 12 Jk−1 rk for all k sufficiently large, and hence that for any
(Hint: Show that Jk+1
constant > 0, the sequence {xk } satisfies xk − x ∗ ≤ for all k sufficiently large.)
✐
11.10 Consider the following modification of our example of failure of continuation
methods:
r (x) x 2 − 1,
a 12 .
Show that for this example there is a zero path for H (x, λ) λ(x 2 − 1) + (1 − λ)(x − a)
that connects ( 12 , 0) to (1, 1), so that continuation methods should work for this choice of
starting point.
303
This is pag
Printer: O
12
CHAPTER
Theory of
Constrained
Optimization
The second part of this book is about minimizing functions subject to constraints on the
variables. A general formulation for these problems is
minn f (x)
x∈IR
subject to
ci (x) 0,
i ∈ E,
ci (x) ≥ 0,
i ∈ I,
(12.1)
where f and the functions ci are all smooth, real-valued functions on a subset of IRn , and
I and E are two finite sets of indices. As before, we call f the objective function, while ci ,
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
i ∈ E are the equality constraints and ci , i ∈ I are the inequality constraints. We define the
feasible set to be the set of points x that satisfy the constraints; that is,
{x | ci (x) 0, i ∈ E; ci (x) ≥ 0, i ∈ I},
(12.2)
so that we can rewrite (12.1) more compactly as
min f (x).
x∈
(12.3)
In this chapter we derive mathematical characterizations of the solutions of (12.3). As
in the unconstrained case, we discuss optimality conditions of two types. Necessary conditions are conditions that must be satisfied by any solution point (under certain assumptions).
Sufficient conditions are those that, if satisfied at a certain point x ∗ , guarantee that x ∗ is in
fact a solution.
For the unconstrained optimization problem of Chapter 2, the optimality conditions
were as follows:
Necessary conditions: Local unconstrained minimizers have ∇ f (x ∗ ) 0 and
∇ 2 f (x ∗ ) positive semidefinite.
Sufficient conditions: Any point x ∗ at which ∇ f (x ∗ ) 0 and ∇ 2 f (x ∗ ) is positive
definite is a strong local minimizer of f .
In this chapter, we derive analogous conditions to characterize the solutions of constrained
optimization problems.
LOCAL AND GLOBAL SOLUTIONS
We have seen already that global solutions are difficult to find even when there are
no constraints. The situation may be improved when we add constraints, since the feasible
set might exclude many of the local minima and it may be comparatively easy to pick the
global minimum from those that remain. However, constraints can also make things more
difficult. As an example, consider the problem
min (x2 + 100)2 + 0.01x12 ,
subject to x2 − cos x1 ≥ 0,
(12.4)
illustrated in Figure 12.1. Without the constraint, the problem has the unique solution
(0, −100)T . With the constraint, there are local solutions near the points
x (k) (kπ, −1)T ,
for k ±1, ±3, ±5, . . ..
Definitions of the different types of local solutions are simple extensions of the corresponding definitions for the unconstrained case, except that now we restrict consideration
to the feasible points in the neighborhood of x ∗ . We have the following definition.
305
306
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
contours of f
x (2)
x (1)
x (3)
x (4)
x (5)
feasible region
Figure 12.1
Constrained problem with many isolated local solutions.
A vector x ∗ is a local solution of the problem (12.3) if x ∗ ∈ and there is a
neighborhood N of x ∗ such that f (x) ≥ f (x ∗ ) for x ∈ N ∩ .
Similarly, we can make the following definitions:
A vector x ∗ is a strict local solution (also called a strong local solution) if x ∗ ∈ and there
is a neighborhood N of x ∗ such that f (x) > f (x ∗ ) for all x ∈ N ∩ with x x ∗ .
A point x ∗ is an isolated local solution if x ∗ ∈ and there is a neighborhood N of x ∗
such that x ∗ is the only local solution in N ∩ .
Note that isolated local solutions are strict, but that the reverse is not true (see
Exercise 12.2).
SMOOTHNESS
Smoothness of objective functions and constraints is an important issue in characterizing solutions, just as in the unconstrained case. It ensures that the objective function and
the constraints all behave in a reasonably predictable way and therefore allows algorithms
to make good choices for search directions.
We saw in Chapter 2 that graphs of nonsmooth functions contain “kinks" or “jumps”
where the smoothness breaks down. If we plot the feasible region for any given constrained
optimization problem, we usually observe many kinks and sharp edges. Does this mean that
the constraint functions that describe these regions are nonsmooth? The answer is often
no, because the nonsmooth boundaries can often be described by a collection of smooth
constraint functions. Figure 12.2 shows a diamond-shaped feasible region in IR2 that could
be described by the single nonsmooth constraint
x1 |x1 | + |x2 | ≤ 1.
(12.5)
It can also be described by the following set of smooth (in fact, linear) constraints:
x1 + x2 ≤ 1,
x1 − x2 ≤ 1,
−x1 + x2 ≤ 1,
−x1 − x2 ≤ 1.
(12.6)
Each of the four constraints represents one edge of the feasible polytope. In general, the constraint functions are chosen so that each one represents a smooth piece of the boundary of .
12.1.
EXAMPLES
Figure 12.2
A feasible region with a
nonsmooth boundary can be
described by smooth constraints.
Nonsmooth, unconstrained optimization problems can sometimes be reformulated as
smooth constrained problems. An example is the unconstrained minimization of a function
f (x) max(x 2 , x),
(12.7)
which has kinks at x 0 and x 1, and the solution at x ∗ 0. We obtain a smooth,
constrained formulation of this problem by adding an artificial variable t and writing
min t
s.t.
t ≥ x,
t ≥ x 2.
(12.8)
Reformulation techniques such as (12.6) and (12.8) are used often in cases where f is a
maximum of a collection of functions or when f is a 1-norm or ∞-norm of a vector
function.
In the examples above we expressed inequality constraints in a slightly different way
from the form ci (x) ≥ 0 that appears in the definition (12.1). However, any collection of
inequality constraints with ≥ and ≤ and nonzero right-hand-sides can be expressed in the
form ci (x) ≥ 0 by simple rearrangement of the inequality.
12.1 EXAMPLES
To introduce the basic principles behind the characterization of solutions of constrained
optimization problems, we work through three simple examples. The discussion here is
informal; the ideas introduced will be made rigorous in the sections that follow.
We start by noting one important item of terminology that recurs throughout the rest
of the book.
307
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
Definition 12.1.
The active set A(x) at any feasible x consists of the equality constraint indices from E
together with the indices of the inequality constraints i for which ci (x) 0; that is,
A(x) E ∪ {i ∈ I | ci (x) 0}.
At a feasible point x, the inequality constraint i ∈ I is said to be active if ci (x) 0
and inactive if the strict inequality ci (x) > 0 is satisfied.
A SINGLE EQUALITY CONSTRAINT
❏ EXAMPLE 12.1
Our first example is a two-variable problem with a single equality constraint:
min x1 + x2
x12 + x22 − 2 0
s.t.
(12.9)
(see Figure 12.3). In the language of (12.1), we have f (x) x1 + x2 , I ∅, E {1}, and
2
c1 (x) x12 + x√
2 − 2. We can see by inspection that the feasible set for this problem is the
circle of radius 2 centered at the origin—just the boundary of this circle, not its interior.
The solution x ∗ is obviously (−1, −1)T . From any other point on the circle, it is easy to
find a way to move that stays feasible
√ (that is, remains on the circle) while decreasing f .
For instance, from the point x ( 2, 0)T any move in the clockwise direction around the
circle has the desired effect.
x2
c1
∆
c1
∆
f
x1
∆
x*
c1
f
f
∆
∆
∆
∆
308
f
Figure 12.3
Problem (12.9), showing
constraint and function
gradients at various feasible
points.
12.1.
EXAMPLES
We also see from Figure 12.3 that at the solution x ∗ , the constraint normal ∇c1 (x ∗ ) is
parallel to ∇ f (x ∗ ). That is, there is a scalar λ∗1 (in this case λ∗1 −1/2) such that
∇ f (x ∗ ) λ∗1 ∇c1 (x ∗ ).
(12.10)
❐
We can derive (12.10) by examining first-order Taylor series approximations to the
objective and constraint functions. To retain feasibility with respect to the function c1 (x) 0, we require any small (but nonzero) step s to satisfy that c1 (x + s) 0; that is,
0 c1 (x + s) ≈ c1 (x) + ∇c1 (x)T s ∇c1 (x)T s.
(12.11)
Hence, the step s retains feasibility with respect to c1 , to first order, when it satisfies
∇c1 (x)T s 0.
(12.12)
Similarly, if we want s to produce a decrease in f , we would have so that
0 > f (x + s) − f (x) ≈ ∇ f (x)T s,
or, to first order,
∇ f (x)T s < 0.
(12.13)
Existence of a small step s that satisfies both (12.12) and (12.13) strongly suggests existence
of a direction d (where the size of d is not small; we could have d ≈ s/s to ensure that
the norm of d is close to 1) with the same properties, namely
∇c1 (x)T d 0 and ∇ f (x)T d < 0.
(12.14)
If, on the other hand, there is no direction d with the properties (12.14), then is it likely that
we cannot find a small step s with the properties (12.12) and (12.13). In this case, x ∗ would
appear to be a local minimizer.
By drawing a picture, the reader can check that the only way that a d satisfying (12.14)
does not exist is if ∇ f (x) and ∇c1 (x) are parallel, that is, if the condition ∇ f (x) λ1 ∇c1 (x)
holds at x, for some scalar λ1 . If in fact ∇ f (x) and ∇c1 (x) are not parallel, we can set
∇c1 (x)∇c1 (x)T
∇ f (x);
d̄ − I −
∇c1 (x)2
It is easy to verify that this d satisfies (12.14).
d
d̄
.
d̄
(12.15)
309
310
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
By introducing the Lagrangian function
L(x, λ1 ) f (x) − λ1 c1 (x),
(12.16)
and noting that ∇x L(x, λ1 ) ∇ f (x) − λ1 ∇c1 (x), we can state the condition (12.10)
equivalently as follows: At the solution x ∗ , there is a scalar λ∗1 such that
∇x L(x ∗ , λ∗1 ) 0.
(12.17)
This observation suggests that we can search for solutions of the equality-constrained
problem (12.9) by seeking stationary points of the Lagrangian function. The scalar quantity
λ1 in (12.16) is called a Lagrange multiplier for the constraint c1 (x) 0.
Though the condition (12.10) (equivalently, (12.17)) appears to be necessary for
an optimal solution of the problem (12.9), it is clearly not sufficient. For instance, in
Example 12.1, condition (12.10) is satisfied at the point x (1, 1)T (with λ1 12 ), but
this point is obviously not a solution—in fact, it maximizes the function f on the circle.
Moreover, in the case of equality-constrained problems, we cannot turn the condition
(12.10) into a sufficient condition simply by placing some restriction on the sign of λ1 . To
see this, consider replacing the constraint x12 + x22 − 2 0 by its negative 2 − x12 − x22 0 in
Example 12.1. The solution of the problem is not affected, but the value of λ∗1 that satisfies
the condition (12.10) changes from λ∗1 − 12 to λ∗1 12 .
A SINGLE INEQUALITY CONSTRAINT
❏ EXAMPLE 12.2
This is a slight modification of Example 12.1, in which the equality constraint is
replaced by an inequality. Consider
min x1 + x2
s.t.
2 − x12 − x22 ≥ 0,
(12.18)
for which the feasible region consists of the circle of problem (12.9) and its interior (see
Figure 12.4). Note that the constraint normal ∇c1 points toward the interior of the feasible
region at each point on the boundary of the circle. By inspection, we see that the solution
is still (−1, −1)T and that the condition (12.10) holds for the value λ∗1 12 . However,
this inequality-constrained problem differs from the equality-constrained problem (12.9)
of Example 12.1 in that the sign of the Lagrange multiplier plays a significant role, as we
now argue.
❐
12.1.
EXAMPLES
As before, we conjecture that a given feasible point x is not optimal if we can find a
small step s that both retains feasibility and decreases the objective function f to first order.
The main difference between problems (12.9) and (12.18) comes in the handling of the
feasibility condition. As in (12.13), the step s improves the objective function, to first order,
if ∇ f (x)T s < 0. Meanwhile, s retains feasibility if
0 ≤ c1 (x + s) ≈ c1 (x) + ∇c1 (x)T s,
so, to first order, feasibility is retained if
c1 (x) + ∇c1 (x)T s ≥ 0.
(12.19)
In determining whether a step s exists that satisfies both (12.13) and (12.19), we
consider the following two cases, which are illustrated in Figure 12.4.
Case I: Consider first the case in which x lies strictly inside the circle, so that the strict
inequality c1 (x) > 0 holds. In this case, any step vector s satisfies the condition (12.19),
provided only that its length is sufficiently small. In fact, whenever ∇ f (x) 0, we can
obtain a step s that satisfies both (12.13) and (12.19) by setting
s −α∇ f (x),
f
∆
x
∆
∆
c1
f
s
x
s
Figure 12.4 Improvement directions s from two feasible points x for the problem
(12.18) at which the constraint is active and inactive, respectively.
311
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
for any positive scalar α sufficiently small. However, this definition does not give a step s
with the required properties when
∇ f (x) 0,
(12.20)
Case II: Consider now the case in which x lies on the boundary of the circle, so that
c1 (x) 0. The conditions (12.13) and (12.19) therefore become
∇ f (x)T s < 0,
∇c1 (x)T s ≥ 0.
The first of these conditions defines an open half-space, while the second defines a closed
half-space, as illustrated in Figure 12.5. It is clear from this figure that the intersection of
these two regions is empty only when ∇ f (x) and ∇c1 (x) point in the same direction, that
is, when
∇ f (x) λ1 ∇c1 (x),
for some λ1 ≥ 0.
(12.21)
Note that the sign of the multiplier is significant here. If (12.10) were satisfied with a negative
value of λ1 , then ∇ f (x) and ∇c1 (x) would point in opposite directions, and we see from
Figure 12.5 that the set of directions that satisfy both (12.13) and (12.19) would make up an
entire open half-plane.
∆
∆
312
f
c1
Any d in this cone is a good search
direction, to first order
Figure 12.5 A direction d that satisfies both (12.13) and (12.19) lies in the
intersection of a closed half-plane and an open half-plane.
12.1.
EXAMPLES
The optimality conditions for both cases I and II can again be summarized neatly
with reference to the Lagrangian function L defined in (12.16). When no first-order feasible
descent direction exists at some point x ∗ , we have that
∇x L(x ∗ , λ∗1 ) 0,
for some λ∗1 ≥ 0,
(12.22)
where we also require that
λ∗1 c1 (x ∗ ) 0.
(12.23)
Condition (12.23) is known as a complementarity condition; it implies that the Lagrange
multiplier λ1 can be strictly positive only when the corresponding constraint c1 is active.
Conditions of this type play a central role in constrained optimization, as we see in the
sections that follow. In case I, we have that c1 (x ∗ ) > 0, so (12.23) requires that λ∗1 0.
Hence, (12.22) reduces to ∇ f (x ∗ ) 0, as required by (12.20). In case II, (12.23) allows λ∗1
to take on a nonnegative value, so (12.22) becomes equivalent to (12.21).
TWO INEQUALITY CONSTRAINTS
❏ EXAMPLE 12.3
Suppose we add an extra constraint to the problem (12.18) to obtain
min x1 + x2
s.t.
2 − x12 − x22 ≥ 0, x2 ≥ 0,
(12.24)
for which the feasible√
region is the half-disk illustrated in Figure 12.6. It is easy to see that
the solution lies at (− 2, 0)T , a point at which both constraints are active. By repeating the
arguments for the previous examples, we would expect a direction d of first-order feasible
descent to satisfy
∇ci (x)T d ≥ 0,
i ∈ I {1, 2},
∇ f (x)T d < 0.
(12.25)
√
However, it is clear from Figure 12.6 that no such direction can exist when x (− 2, 0)T .
The conditions ∇ci (x)T d ≥ 0, i 1, 2, are both satisfied only if d lies in the quadrant
defined by ∇c1 (x) and ∇c2 (x), but it is clear by inspection that all vectors d in this quadrant
satisfy ∇ f (x)T d ≥ 0.
Let us see how the
√ Lagrangian and its derivatives behave for the problem (12.24) and
the solution point (− 2, 0)T . First, we include an additional term λi ci (x) in the Lagrangian
for each additional constraint, so the definition of L becomes
L(x, λ) f (x) − λ1 c1 (x) − λ2 c2 (x),
313
CHAPTER 12.
∆
THEORY
OF
CONSTRAINED OPTIMIZATION
c2
∆
f
∆
314
Figure 12.6
Problem (12.24), illustrating the
gradients of the active constraints
and objective at the solution.
c1
where λ (λ1 , λ2 )T is the vector of Lagrange multipliers. The extension of condition
(12.22) to this case is
∇x L(x ∗ , λ∗ ) 0,
for some λ∗ ≥ 0,
(12.26)
where the inequality λ∗ ≥ 0 means that all components of λ∗ are required to be nonnegative.
By applying the complementarity condition (12.23) to both inequality constraints, we obtain
λ∗1 c1 (x ∗ ) 0,
λ∗2 c2 (x ∗ ) 0.
(12.27)
√
When x ∗ (− 2, 0)T , we have
∇ f (x ∗ ) 1
1
,
∇c1 (x ∗ ) √ 2 2
0
,
∇c2 (x ∗ ) 0
1
,
so that it is easy to verify that ∇x L(x ∗ , λ∗ ) 0 when we select λ∗ as follows:
∗
λ √ 1/(2 2)
1
.
Note that both components of λ∗ are positive, so that (12.26) is satisfied.
We consider now some other feasible points that are not solutions of (12.24), and
examine the properties of the
√Lagrangian and its gradient at these points.
For the point x ( 2, 0)T , we again have that both constraints are active (see
Figure 12.7). However, it s easy to identify vectors d that satisfies (12.25): d (−1, 0)T
is one such vector (there are many others). For this value of √
x it is easy to verify that the
condition ∇x L(x, λ) 0 is satisfied only when λ (−1/(2 2), 1)T . Note that the first
component λ1 is negative, so that the conditions (12.26) are not satisfied at this point.
Finally, we consider the point x (1, 0)T , at which only the second constraint c2 is
active. Since any small step s away from this point will continue to satisfy c1 (x + s) > 0, we
need to consider only the behavior of c2 and f in determining whether s is indeed a feasible
12.2.
TANGENT CONE
AND
CONSTRAINT QUALIFICATIONS
c2
f
∆
∆
∆
c1
Figure 12.7
Problem (12.24), illustrating
the gradients of the active
constraints and objective at a
nonoptimal point.
descent step. Using the same reasoning as in the earlier examples, we find that the direction
of feasible descent d must satisfy
∇c2 (x)T d ≥ 0,
∇ f (x)T d < 0.
(12.28)
By noting that
∇ f (x) 1
1
,
∇c2 (x) 0
1
,
T
it is easy to verify that the vector d − 12 , 14 satisfies (12.28) and is therefore a descent
direction.
To show that optimality conditions (12.26) and (12.27) fail, we note first from (12.27)
that since c1 (x) > 0, we must have λ1 0. Therefore, in trying to satisfy ∇x L(x, λ) 0,
we are left to search for a value λ2 such that ∇ f (x) − λ2 ∇c2 (x) 0. No such λ2 exists, and
thus this point fails to satisfy the optimality conditions.
❐
12.2 TANGENT CONE AND CONSTRAINT QUALIFICATIONS
In this section we define the tangent cone T (x ∗ ) to the closed convex set at a point
x ∗ ∈ , and also the set F(x ∗ ) of first-order feasible directions at x ∗ . We also discuss
constraint qualifications. In the previous section, we determined whether or not it was
possible to take a feasible descent step away from a given feasible point x by examining
the first derivatives of f and the constraint functions ci . We used the first-order Taylor
series expansion of these functions about x to form an approximate problem in which both
objective and constraints are linear. This approach makes sense, however, only when the
linearized approximation captures the essential geometric features of the feasible set near
the point x in question. If, near x, the linearization is fundamentally different from the
315
316
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
feasible set (for instance, it is an entire plane, while the feasible set is a single point) then
we cannot expect the linear approximation to yield useful information about the original
problem. Hence, we need to make assumptions about the nature of the constraints ci that
are active at x to ensure that the linearized approximation is similar to the feasible set, near
x. Constraint qualifications are assumptions that ensure similarity of the constraint set and its linearized approximation, in a neighborhood of x ∗ .
Given a feasible point x, we call {z k } a feasible sequence approaching x if z k ∈ for all
k sufficiently large and z k → x.
Later, we characterize a local solution of (12.1) as a point x at which all feasible
sequences approaching x have the property that f (z k ) ≥ f (x) for all k sufficiently large,
and we will derive practical, verifiable conditions under which this property holds. We lay
the groundwork in this section by characterizing the directions in which we can step away
from x while remaining feasible.
A tangent is a limiting direction of a feasible sequence.
Definition 12.2.
The vector d is said to be a tangent (or tangent vector) to at a point x if there are a
feasible sequence {z k } approaching x and a sequence of positive scalars {tk } with tk → 0 such
that
lim
k→∞
zk − x
d.
tk
(12.29)
The set of all tangents to at x ∗ is called the tangent cone and is denoted by T (x ∗ ).
It is easy to see that the tangent cone is indeed a cone, according to the definition
(A.36). If d is a tangent vector with corresponding sequences {z k } and {tk }, then by replacing
each tk by α −1 tk , for any α > 0, we find that αd ∈ T (x ∗ ) also. We obtain that 0 ∈ T (x)
by setting z k ≡ x in the definition of feasible sequence.
We turn now to the linearized feasible direction set, which we define as follows.
Definition 12.3.
Given a feasible point x and the active constraint set A(x) of Definition 12.1, the set of
linearized feasible directions F(x) is
d T ∇ci (x) 0,
F(x) d T
d ∇ci (x) ≥ 0,
for all i ∈ E,
for all i ∈ A(x) ∩ I
.
As with the tangent cone, it is easy to verify that F(x) is a cone, according to the definition
(A.36).
It is important to note that the definition of tangent cone does not rely on the algebraic
specification of the set , only on its geometry. The linearized feasible direction set does,
however, depend on the definition of the constraint functions ci , i ∈ E ∪ I.
12.2.
TANGENT CONE
AND
CONSTRAINT QUALIFICATIONS
c1
feasible sequence zk
_ f
Figure 12.8
Constraint normal, objective
gradient, and feasible
sequence for problem (12.9).
tangent d
We illustrate the tangent cone and the linearized feasible direction set by revisiting
Examples 12.1 and 12.2.
❏ EXAMPLE 12.4
(EXAMPLE 12.1, REVISITED)
Figure 12.8 shows the problem
(12.9), the equality-constrained problem
in which the
√
√
feasible set is a circle of radius 2, near the nonoptimal point x (− 2, 0)T . The figure
also shows a feasible sequence approaching x. This sequence could be defined analytically
by the formula
zk − 2 − 1/k 2
−1/k
.
(12.30)
By choosing tk z k − x, we find that d (0, −1)T is a tangent. Note that the objective
function f (x) x1 + x2 increases as we move along the sequence (12.30); in fact, we have
f (z k+1 ) > f (z k ) for all k 2, 3, . . .. It follows that f (z k ) < f (x) for k 2, 3, . . ., so x
cannot be a solution of (12.9).
√
Another feasible sequence is one that approaches x (− 2, 0)T from the opposite
direction. Its elements are defined by
zk − 2 − 1/k 2
1/k
.
It is easy to show that f decreases along this sequence and that the tangents corresponding
√
to this sequence are d (0, α)T . In summary, the tangent cone at x (− 2, 0)T is
{(0, d2 )T | d2 ∈ IR}.
317
318
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
For the definition (12.9) of this set, and Definition 12.3, we have that d (d1 , d2 )T ∈
F(x) if
0 ∇c1 (x) d T
2x1
T 2x2
d1
d2
√
−2 2d1 .
Therefore, we obtain F(x) {(0, d2 )T | d2 ∈ IR}. In this case, we have T (x) F(x).
Suppose that the feasible set is defined instead by the formula
{x | c1 (x) 0}, where c1 (x) (x12 + x22 − 2)2 0.
(12.31)
(Note that is the same, but its algebraic specification has changed.) The vector d belongs
to the linearized feasible set if
0 ∇c1 (x)T d 4(x12 + x22 − 2)x1
T 4(x12 + x22 − 2)x2
d1
d2
0
0
T d1
d2
,
which is true for all (d1 , d2 )T . Hence, we have F(x) IR2 , so for this algebraic specification
of , the tangent cone and linearized feasible sets differ.
❐
❏ EXAMPLE 12.5
(EXAMPLE 12.2, REVISITED)
We now reconsider problem (12.18) in Example 12.2. The solution x (−1, −1)T is
the same as in the equality-constrained case, but there is a much more extensive collection
of feasible sequences that converge to any given feasible point (see Figure 12.9).
x2
x1
Figure 12.9
Feasible sequences converging to a particular
feasible point for the region defined by
x12 + x22 ≤ 2.
12.2.
TANGENT CONE
AND
CONSTRAINT QUALIFICATIONS
√
From the point x (− 2, 0)T , the various feasible sequences defined above for the
equality-constrained problem are still feasible
√ for (12.18). There are also infinitely many
feasible sequences that converge to x (− 2, 0)T along a straight line from the interior of
the circle. These sequences have the form
√
z k (− 2, 0)T + (1/k)w,
where w is any vector √
whose first component is positive (w1 > 0). The point z k is feasible
provided that z k ≤ 2, that is,
√
(− 2 + w1 /k)2 + (w2 /k)2 ≤ 2,
√
which is true when k ≥ (w12 + w22 )/(2 2w1 ). In addition to these straight-line √
feasible
sequences, we can also define an infinite variety of sequences that approach (− 2, 0)T
along
√ a curve from the interior of the circle. To summarize, the tangent cone to this set at
(− 2, 0)T is {(w1 , w2 )T | w1 ≥ 0}.
For the definition (12.18) of this feasible set, we have from Definition 12.3 that
d ∈ F(x) if
0 ≤ ∇c1 (x) d T
−2x1
T −2x2
d1
d2
√
2 2d1 .
Hence, we obtain F(x) T (x) for this particular algebraic specification of the feasible
set.
❐
Constraint qualifications are conditions under which the linearized feasible set F(x)
is similar to the tangent cone T (x). In fact, most constraint qualifications ensure that these
two sets are identical. As mentioned earlier, these conditions ensure that the F(x), which is
constructed by linearizing the algebraic description of the set at x, captures the essential
geometric features of the set in the vicinity of x, as represented by T (x).
Revisiting Example 12.4, we see that both T (x) and F(x) consist of the vertical axis,
which is qualitatively similar to the set − {x} in the neighborhood of x. As a further
example, consider the constraints
c1 (x) 1 − x12 − (x2 − 1)2 ≥ 0,
c2 (x) −x2 ≥ 0,
(12.32)
for which the feasible set is the single point {(0, 0)T } (see Figure 12.10). For this point
x (0, 0)T , it is obvious that that tangent cone is T (x) {(0, 0)T }, since all feasible
sequences approaching x must have z k x (0, 0)T for all k sufficiently large. Moreover,
it is easy to show that linearized approximation to the feasible set F(x) is
F(x ∗ ) {(d1 , 0)T | d1 ∈ IR},
319
320
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
Figure 12.10 Problem (12.32), for which the feasible set is the single point of
intersection between circle and line.
that is, the entire horizontal axis. In this case, the linearized feasible direction set does not
capture the geometry of the feasible set, so constraint qualifications are not satisfied.
The constraint qualification most often used in the design of algorithms is the subject
of the next definition.
Definition 12.4 (LICQ).
Given the point x and the active set A(x) defined in Definition 12.1, we say that the linear
independence constraint qualification (LICQ) holds if the set of active constraint gradients
{∇ci (x), i ∈ A(x)} is linearly independent.
Note that this condition is not satisfied for the examples (12.32) and (12.31). In general, if
LICQ holds, none of the active constraint gradients can be zero. We mention other constraint
qualifications in Section 12.6.
12.3 FIRST-ORDER OPTIMALITY CONDITIONS
In this section, we state first-order necessary conditions for x ∗ to be a local minimizer
and show how these conditions are satisfied on a small example. The proof of the result is
presented in subsequent sections.
As a preliminary to stating the necessary conditions, we define the Lagrangian function
for the general problem (12.1).
L(x, λ) f (x) −
λi ci (x).
(12.33)
i∈E∪I
(We had previously defined special cases of this function for the examples of Section 12.1.)
12.3.
FIRST-ORDER OPTIMALITY CONDITIONS
The necessary conditions defined in the following theorem are called first-order conditions because they are concerned with properties of the gradients (first-derivative vectors)
of the objective and constraint functions. These conditions are the foundation for many of
the algorithms described in the remaining chapters of the book.
Theorem 12.1 (First-Order Necessary Conditions).
Suppose that x ∗ is a local solution of (12.1), that the functions f and ci in (12.1) are
continuously differentiable, and that the LICQ holds at x ∗ . Then there is a Lagrange multiplier
vector λ∗ , with components λi∗ , i ∈ E ∪ I, such that the following conditions are satisfied at
(x ∗ , λ∗ )
∇x L(x ∗ , λ∗ ) 0,
for all i ∈ E,
(12.34b)
∗
for all i ∈ I,
(12.34c)
for all i ∈ I,
(12.34d)
for all i ∈ E ∪ I.
(12.34e)
ci (x ) 0,
ci (x ) ≥ 0,
λi∗
∗
λi∗ ci (x
(12.34a)
∗
≥ 0,
) 0,
The conditions (12.34) are often known as the Karush–Kuhn–Tucker conditions, or
KKT conditions for short. The conditions (12.34e) are complementarity conditions; they
imply that either constraint i is active or λi∗ 0, or possibly both. In particular, the
Lagrange multipliers corresponding to inactive inequality constraints are zero, we can omit
the terms for indices i ∈
/ A(x ∗ ) from (12.34a) and rewrite this condition as
0 ∇x L(x ∗ , λ∗ ) ∇ f (x ∗ ) −
λi∗ ∇ci (x ∗ ).
(12.35)
i∈A(x ∗ )
A special case of complementarity is important and deserves its own definition.
Definition 12.5 (Strict Complementarity).
Given a local solution x ∗ of (12.1) and a vector λ∗ satisfying (12.34), we say that the
strict complementarity condition holds if exactly one of λi∗ and ci (x ∗ ) is zero for each index
i ∈ I. In other words, we have that λi∗ > 0 for each i ∈ I ∩ A(x ∗ ).
Satisfaction of the strict complementarity property usually makes it easier for algorithms to
determine the active set A(x ∗ ) and converge rapidly to the solution x ∗ .
For a given problem (12.1) and solution point x ∗ , there may be many vectors λ∗ for
which the conditions (12.34) are satisfied. When the LICQ holds, however, the optimal λ∗
is unique (see Exercise 12.17).
The proof of Theorem 12.1 is quite complex, but it is important to our understanding
of constrained optimization, so we present it in the next section. First, we illustrate the KKT
conditions with another example.
321
322
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
x2
x1
Inequality-constrained problem (12.36) with solution at (1, 0)T .
Figure 12.11
❏ EXAMPLE 12.6
Consider the feasible region illustrated in Figure 12.2 and described by the four
constraints (12.6). By restating the constraints in the standard form of (12.1) and including
an objective function, the problem becomes
⎡
min
x
3
x1 −
2
2
1
+ x2 −
2
4
s.t.
1 − x1 − x2
⎢
⎢ 1 − x1 + x2
⎢
⎢ 1+x −x
1
2
⎣
1 + x1 + x2
⎤
⎥
⎥
⎥ ≥ 0.
⎥
⎦
(12.36)
It is fairly clear from Figure 12.11 that the solution is x ∗ (1, 0)T . The first and second
constraints in (12.36) are active at this point. Denoting them by c1 and c2 (and the inactive
constraints by c3 and c4 ), we have
⎡
⎤
−1
∇ f (x ∗ ) ⎣
1 ⎦,
−
2
∇c1 (x ∗ ) −1
−1
,
∇c2 (x ∗ ) −1
1
.
Therefore, the KKT conditions (12.34a)–(12.34e) are satisfied when we set
λ∗ 3
4
T
, 14 , 0, 0 .
❐
12.4.
12.4
FIRST-ORDER OPTIMALITY CONDITIONS: PROOF
FIRST-ORDER OPTIMALITY CONDITIONS: PROOF
We now develop a proof of Theorem 12.1. A number of key subsidiary results are required,
so the development is quite long. However, a complete treatment is worthwhile, since these
results are so fundamental to the field of optimization.
RELATING THE TANGENT CONE AND THE FIRST-ORDER FEASIBLE
DIRECTION SET
The following key result uses a constraint qualification (LICQ) to relate the tangent
cone of Definition 12.2 to the set F of first-order feasible directions of Definition 12.3. In
the proof below and in later results, we use the notation A(x ∗ ) to represent the matrix whose
rows are the active constraint gradients at the optimal point, that is,
A(x ∗ )T [∇ci (x ∗ )]i∈A(x ∗ ) ,
(12.37)
where the active set A(x ∗ ) is defined as in Definition 12.1.
Lemma 12.2.
Let x ∗ be a feasible point. The following two statements are true.
(i) T (x ∗ ) ⊂ F(x ∗ ).
(ii) If the LICQ condition is satisfied at x ∗ , then F(x ∗ ) T (x ∗ ).
PROOF. Without loss of generality, let us assume that all the constraints ci (·), i 1, 2, . . . , m, are active at x ∗ . (We can arrive at this convenient ordering by simply dropping all
inactive constraints—which are irrelevant in some neighborhood of x ∗ —and renumbering
the active constraints that remain.)
To prove (i), let {z k } and {tk } be the sequences for which (12.29) is satisfied, that is,
lim
k→∞
zk − x ∗
d.
tk
(Note in particular that tk > 0 for all k.) From this definition, we have that
z k x ∗ + tk d + o(tk ).
By taking i ∈ E and using Taylor’s theorem, we have that
1
ci (z k )
tk
1
ci (x ∗ ) + tk ∇ci (x ∗ )T d + o(tk )
tk
o(tk )
∇ci (x ∗ )T d +
.
tk
0
(12.38)
323
324
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
By taking the limit as k → ∞, the last term in this expression vanishes, and we have
∇ci (x ∗ )T d 0, as required. For the active inequality constraints i ∈ A(x ∗ ) ∩ I, we have
similarly that
1
ci (z k )
tk
1
ci (x ∗ ) + tk ∇ci (x ∗ )T d + o(tk )
tk
o(tk )
.
∇ci (x ∗ )T d +
tk
0≤
Hence, by a similar limiting argument, we have that ∇ci (x ∗ )T d ≥ 0, as required.
For (ii), we use the implicit function theorem (see the Appendix or Lang [187, p. 131]
for a statement of this result). First, since the LICQ holds, we have from Definition 12.4 that
the m × n matrix A(x ∗ ) of active constraint gradients has full row rank m. Let Z be a matrix
whose columns are a basis for the null space of A(x ∗ ); that is,
Z ∈ IRn×(n−m) ,
A(x ∗ )Z 0.
Z has full column rank,
(12.39)
(See the related discussion in Chapter 16.) Choose d ∈ F(x ∗ ) arbitrarily, and suppose that
{tk }∞
k0 is any sequence of positive scalars such limk→∞ tk 0. Define the parametrized
system of equations R : IRn × IR → IRn by
R(z, t) c(z) − t A(x ∗ )d
Z T (z − x ∗ − td)
0
0
.
(12.40)
We claim that the solutions z z k of this system for small t tk > 0 give a feasible sequence
that approaches x ∗ and satisfies the definition (12.29).
At t 0, z x ∗ , and the Jacobian of R at this point is
∇z R(x ∗ , 0) A(x ∗ )
ZT
,
(12.41)
which is nonsingular by construction of Z . Hence, according to the implicit function
theorem, the system (12.40) has a unique solution z k for all values of tk sufficiently small.
Moreover, we have from (12.40) and Definition 12.3 that
i ∈ E ⇒ ci (z k ) tk ∇ci (x ∗ )T d 0,
∗
∗ T
i ∈ A(x ) ∩ I ⇒ ci (z k ) tk ∇ci (x ) d ≥ 0,
so that z k is indeed feasible.
(12.42a)
(12.42b)
12.4.
FIRST-ORDER OPTIMALITY CONDITIONS: PROOF
It remains to verify that (12.29) holds for this choice of {z k }. Using the fact that
R(z k , tk ) 0 for all k together with Taylor’s theorem, we find that
0 R(z k , tk ) c(z k ) − tk A(x ∗ )d
Z T (z k − x ∗ − tk d)
A(x ∗ )(z k − x ∗ ) + o(z k − x ∗ ) − tk A(x ∗ )d
A(x ∗ )
ZT
Z T (z k − x ∗ − tk d)
(z k − x ∗ − tk d) + o(z k − x ∗ ).
By dividing this expression by tk and using nonsingularity of the coefficient matrix in the
first term, we obtain
zk − x ∗
z k − x ∗ ,
d +o
tk
tk
from which it follows that (12.29) is satisfied (for x x ∗ ). Hence, d ∈ T (x ∗ ) for an
arbitrary d ∈ F(x ∗ ), so the proof of (ii) is complete.
A FUNDAMENTAL NECESSARY CONDITION
As mentioned above, a local solution of (12.1) is a point x at which all feasible sequences
have the property that f (z k ) ≥ f (x) for all k sufficiently large. The following result shows
that if such a sequence exists, then its limiting directions must make a nonnegative inner
product with the objective function gradient.
Theorem 12.3.
If x ∗ is a local solution of (12.1), then we have
∇ f (x ∗ )T d ≥ 0, for all d ∈ T (x ∗ ).
(12.43)
PROOF. Suppose for contradiction that there is a tangent d for which ∇ f (x ∗ )T d < 0. Let
{z k } and {tk } be the sequences satisfying Definition 12.2 for this d. We have that
f (z k ) f (x ∗ ) + (z k − x ∗ )T ∇ f (x ∗ ) + o(z k − x ∗ )
f (x ∗ ) + tk d T ∇ f (x ∗ ) + o(tk ),
325
326
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
Figure 12.12
Problem (12.44), showing various
limiting directions of feasible
sequences at the point (0, 0)T .
where the second line follows from (12.38). Since d T ∇ f (x ∗ ) < 0, the remainder term is
eventually dominated by the first-order term, that is,
f (z k ) < f (x ∗ ) + 12 tk d T ∇ f (x ∗ ),
for all k sufficiently large.
Hence, given any open neighborhood of x ∗ , we can choose k sufficiently large that z k lies
within this neighborhood and has a lower value of the objective f . Therefore, x ∗ is not a
local solution.
The converse of this result is not necessarily true. That is, we may have ∇ f (x ∗ )T d ≥ 0
for all d ∈ T (x ∗ ), yet x ∗ is not a local minimizer. An example is the following problem in
two unknowns, illustrated in Figure 12.12
min x2
subject to x2 ≥ −x12 .
(12.44)
This problem is actually unbounded, but let us examine its behavior at x ∗ (0, 0)T . It is
not difficult to show that all limiting directions d of feasible sequences must have d2 ≥ 0, so
that ∇ f (x ∗ )T d d2 ≥ 0. However, x ∗ is clearly not a local minimizer; the point (α, −α 2 )T
for α > 0 has a smaller function value than x ∗ , and can be brought arbitrarily close to x ∗
by setting α sufficiently small.
FARKAS’ LEMMA
The most important step in proving Theorem 12.1 is a classical theorem of the
alternative known as Farkas’ Lemma. This lemma considers a cone K defined as follows:
K {By + Cw | y ≥ 0},
(12.45)
12.4.
FIRST-ORDER OPTIMALITY CONDITIONS: PROOF
b1
g
b1
g
b2
b2
b3
b3
d
Figure 12.13
(right).
Farkas’ Lemma: Either g ∈ K (left) or there is a separating hyperplane
where B and C are matrices of dimension n × m and n × p, respectively, and y and w are
vectors of appropriate dimensions. Given a vector g ∈ IRn , Farkas’ Lemma states that one
(and only one) of two alternatives is true. Either g ∈ K , or else there is a vector d ∈ IRn such
that
g T d < 0,
B T d ≥ 0,
C T d 0.
(12.46)
The two cases are illustrated in Figure 12.13 for the case of B with three columns, C null,
and n 2. Note that in the second case, the vector d defines a separating hyperplane, which
is a plane in IRn that separates the vector g from the cone K .
Lemma 12.4 (Farkas).
Let the cone K be defined as in (12.45). Given any vector g ∈ IRn , we have either that
g ∈ K or that there exists d ∈ IRn satisfying (12.46), but not both.
PROOF. We show first that the two alternatives cannot hold simultaneously. If g ∈ K , there
exist vectors y ≥ 0 and w such that g By + Cw. If there also exists a d with the property
(12.46), we have by taking inner products that
0 > d T g d T By + d T Cw (B T d)T y + (C T d)T w ≥ 0,
where the final inequality follows from C T d 0, B T d ≥ 0, and y ≥ 0. Hence, we cannot
have both alternatives holding at once.
We now show that one of the alternatives holds. To be precise, we show how to
construct d with the properties (12.46) in the case that g ∈
/ K . For this part of the proof,
we need to use the property that K is a closed set—a fact that is intuitively obvious but not
trivial to prove (see Lemma 12.15 in the Notes and References below). Let ŝ be the vector
327
328
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
in K that is closest to g in the sense of the Euclidean norm. Because K is closed, ŝ is well
defined and is given by the solution of the following optimization problem:
min s − g22
subject to s ∈ K .
(12.47)
Since ŝ ∈ K , we have from the fact that K is a cone that α ŝ ∈ K for all scalars α ≥ 0. Since
α ŝ − g22 is minimized by α 1, we have by simple calculus that
d
2
0 ⇒ −2ŝ T g + 2t ŝ T ŝ α1 0
α ŝ − g2 dα
α1
⇒ ŝ T (ŝ − g) 0.
(12.48)
Now, let s be any other vector in K . Since K is convex, we have by the minimizing property
of ŝ that
ŝ + θ (s − ŝ) − g22 ≥ ŝ − g22
for all θ ∈ [0, 1],
and hence
2θ (s − ŝ)T (ŝ − g) + θ 2 s − ŝ22 ≥ 0.
By dividing this expression by θ and taking the limit as θ ↓ 0, we have (s − ŝ)T (ŝ − g) ≥ 0.
Therefore, because of (12.48),
s T (ŝ − g) ≥ 0,
for all s ∈ K .
(12.49)
We claim now that the vector
d ŝ − g
satisfies the conditions (12.46). Note that d 0 because g ∈
/ K . We have from (12.48) that
d T g d T (ŝ − d) (ŝ − g)T ŝ − d T d −d22 < 0,
so that d satisfies the first property in (12.46).
From (12.49), we have that d T s ≥ 0 for all s ∈ K , so that
d T (By + Cw) ≥ 0 for all y ≥ 0 and all w.
By fixing y 0 we have that (C T d)T w ≥ 0 for all w, which is true only if C T d 0. By
fixing w 0, we have that (B T d)T y ≥ 0 for all y ≥ 0, which is true only if B T d ≥ 0. Hence,
d also satisfies the second and third properties in (12.46) and our proof is complete.
12.4.
FIRST-ORDER OPTIMALITY CONDITIONS: PROOF
By applying Lemma 12.4 to the cone N defined by
N
⎧
⎨ ⎩
λi ∇ci (x ∗ ),
λi ≥ 0 for i ∈ A(x ∗ ) ∩ I
i∈A(x ∗ )
⎫
⎬
⎭
,
(12.50)
and setting g ∇ f (x ∗ ), we have that either
∇ f (x ∗ ) λi ∇ci (x ∗ ) A(x ∗ )T λ∗ ,
λi ≥ 0 for i ∈ A(x ∗ ) ∩ I,
(12.51)
i∈A(x ∗ )
or else there is a direction d such that d T ∇ f (x ∗ ) < 0 and d ∈ F(x ∗ ).
PROOF OF THEOREM 12.1
Lemmas 12.2 and 12.4 can be combined to give the KKT conditions described in
Theorem 12.1. We work through the final steps of the proof here. Suppose that x ∗ ∈ IRn is a
feasible point at which the LICQ holds. The theorem claims that if x ∗ is a local solution for
(12.1), then there is a vector λ∗ ∈ IRm that satisfies the conditions (12.34).
We show first that there are multipliers λi , i ∈ A(x ∗ ), such that (12.51) is satisfied.
Theorem 12.3 tells us that d T ∇ f (x ∗ ) ≥ 0 for all tangent vectors d ∈ T (x ∗ ). From
Lemma 12.2, since LICQ holds, we have that T (x ∗ ) F(x ∗ ). By putting these two
statements together, we find that d T ∇ f (x ∗ ) ≥ 0 for all d ∈ F(x ∗ ). Hence, from Lemma 12.4,
there is a vector λ for which (12.51) holds, as claimed.
We now define the vector λ∗ by
λi∗
λi ,
i ∈ A(x ∗ ),
0,
i ∈ I\A(x ∗ ),
(12.52)
and show that this choice of λ∗ , together with our local solution x ∗ , satisfies the conditions
(12.34). We check these conditions in turn.
• The condition (12.34a) follows immediately from (12.51) and the definitions (12.33)
of the Lagrangian function and (12.52) of λ∗ .
• Since x ∗ is feasible, the conditions (12.34b) and (12.34c) are satisfied.
• We have from (12.51) that λi∗ ≥ 0 for i ∈ A(x ∗ ) ∩ I, while from (12.52), λi∗ 0 for
i ∈ I\A(x ∗ ). Hence, λi∗ ≥ 0 for i ∈ I, so that (12.34d) holds.
• We have for i ∈ A(x ∗ ) ∩ I that ci (x ∗ ) 0, while for i ∈ I\A(x ∗ ), we have λi∗ 0.
Hence λi∗ ci (x ∗ ) 0 for i ∈ I, so that (12.34e) is satisfied as well.
The proof is complete.
329
330
CHAPTER 12.
12.5
THEORY
OF
CONSTRAINED OPTIMIZATION
SECOND-ORDER CONDITIONS
So far, we have described first-order conditions—the KKT conditions—which tell us how
the first derivatives of f and the active constraints ci are related to each other at a solution x ∗ .
When these conditions are satisfied, a move along any vector w from F(x ∗ ) either increases
the first-order approximation to the objective function (that is, w T ∇ f (x ∗ ) > 0), or else
keeps this value the same (that is, w T ∇ f (x ∗ ) 0).
What role do the second derivatives of f and the constraints ci play in optimality
conditions? We see in this section that second derivatives play a “tiebreaking” role. For
the directions w ∈ F(x ∗ ) for which w T ∇ f (x ∗ ) 0, we cannot determine from first
derivative information alone whether a move along this direction will increase or decrease
the objective function f . Second-order conditions examine the second derivative terms in
the Taylor series expansions of f and ci , to see whether this extra information resolves the
issue of increase or decrease in f . Essentially, the second-order conditions concern the
curvature of the Lagrangian function in the “undecided” directions—the directions w ∈
F(x ∗ ) for which w T ∇ f (x ∗ ) 0.
Since we are discussing second derivatives, stronger smoothness assumptions are
needed here than in the previous sections. For the purpose of this section, f and ci ,
i ∈ E ∪ I, are all assumed to be twice continuously differentiable.
Given F(x ∗ ) from Definition 12.3 and some Lagrange multiplier vector λ∗ satisfying
the KKT conditions (12.34), we define the critical cone C(x ∗ , λ∗ ) as follows:
C(x ∗ , λ∗ ) {w ∈ F(x ∗ ) | ∇ci (x ∗ )T w 0, all i ∈ A(x ∗ ) ∩ I with λi∗ > 0}.
Equivalently,
⎧
∗ T
⎪
⎨ ∇ci (x ) w 0,
w ∈ C(x ∗ , λ∗ ) ⇔
∇ci (x ∗ )T w 0,
⎪
⎩
∇ci (x ∗ )T w ≥ 0,
for all i ∈ E,
for all i ∈ A(x ∗ ) ∩ I with λi∗ > 0,
∗
for all i ∈ A(x ) ∩ I with
λi∗
(12.53)
0.
The critical cone contains those directions w that would tend to “adhere” to the active
inequality constraints even when we were to make small changes to the objective (those
indices i ∈ I for which the Lagrange multiplier component λi∗ is positive), as well as to the
equality constraints. From the definition (12.53) and the fact that λi∗ 0 for all inactive
components i ∈ I\A(x ∗ ), it follows immediately that
w ∈ C(x ∗ , λ∗ ) ⇒ λi∗ ∇ci (x ∗ )T w 0 for all i ∈ E ∪ I.
(12.54)
Hence, from the first KKT condition (12.34a) and the definition (12.33) of the Lagrangian
function, we have that
λi∗ w T ∇ci (x ∗ ) 0.
(12.55)
w ∈ C(x ∗ , λ∗ ) ⇒ w T ∇ f (x ∗ ) i∈E∪I
12.5.
SECOND-ORDER CONDITIONS
x2
F
C
∆
− f
x1
Figure 12.14
Problem (12.56), showing
F(x ∗ ) and C(x ∗ , λ∗ ).
Hence the critical cone C(x ∗ , λ∗ ) contains directions from F(x ∗ ) for which it is not clear
from first derivative information alone whether f will increase or decrease.
❏ EXAMPLE 12.7
Consider the problem
min x1
subject to x2 ≥ 0, 1 − (x1 − 1)2 − x22 ≥ 0,
(12.56)
illustrated in Figure 12.14. It is not difficult to see that the solution is x ∗ (0, 0)T , with
active set A(x ∗ ) {1, 2} and a unique optimal Lagrange multiplier λ∗ (0, 0.5)T . Since
the gradients of the active constraints at x ∗ are (0, 1)T and (2, 0)T , respectively, the LICQ
holds, so the optimal multiplier is unique. The linearized feasible set is then
F(x ∗ ) {d | d ≥ 0},
while the critical cone is
C(x ∗ , λ∗ ) {(0, w2 )T | w2 ≥ 0}.
❐
331
332
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
The first theorem defines a necessary condition involving the second derivatives: If
x ∗ is a local solution, then the Hessian of the Lagrangian has nonnegative curvature along
critical directions (that is, the directions in C(x ∗ , λ∗ )).
Theorem 12.5 (Second-Order Necessary Conditions).
Suppose that x ∗ is a local solution of (12.1) and that the LICQ condition is satisfied. Let
∗
λ be the Lagrange multiplier vector for which the KKT conditions (12.34) are satisfied. Then
w T ∇x2x L(x ∗ , λ∗ )w ≥ 0,
for all w ∈ C(x ∗ , λ∗ ).
(12.57)
PROOF. Since x ∗ is a local solution, all feasible sequences {z k } approaching x ∗ must have
f (z k ) ≥ f (x ∗ ) for all k sufficiently large. Our approach in this proof is to construct a
feasible sequence whose limiting direction is w and show that the property f (z k ) ≥ f (x ∗ )
implies that (12.57) holds.
Since w ∈ C(x ∗ , λ∗ ) ⊂ F(x ∗ ), we can use the technique in the proof of Lemma 12.2
to choose a sequence {tk } of positive scalars and to construct a feasible sequence {z k }
approaching x ∗ such that
lim
k→∞
zk − x ∗
w,
tk
(12.58)
which we can write also as (12.58) that
z k − x ∗ tk w + o(tk ).
(12.59)
Because of the construction technique for {z k }, we have from formula (12.42) that
ci (z k ) tk ∇ci (x ∗ )T w,
for all i ∈ A(x ∗ )
(12.60)
From (12.33), (12.60), and (12.54), we have
L(z k , λ∗ ) f (z k ) −
λi∗ ci (z k )
i∈E∪I
f (z k ) − tk
λi∗ ∇ci (x ∗ )T w
i∈A(x ∗ )
f (z k ),
(12.61)
On the other hand, we can perform a Taylor series expansion to obtain an estimate of
L(z k , λ∗ ) near x ∗ . By using Taylor’s theorem expression (2.6) and continuity of the Hessians
∇ 2 f and ∇ 2 ci , i ∈ E ∪ I, we obtain
L(z k , λ∗ ) L(x ∗ , λ∗ ) + (z k − x ∗ )T ∇x L(x ∗ , λ∗ )
+
1
(z
2 k
∗ T
−x )
∇x2x L(x ∗ , λ∗ )(z k
(12.62)
∗
∗ 2
− x ) + o(z k − x ).
12.5.
SECOND-ORDER CONDITIONS
By the complementarity conditions (12.34e), we have L(x ∗ , λ∗ ) f (x ∗ ). From (12.34a),
the second term on the right-hand side is zero. Hence, using (12.59), we can rewrite (12.62)
as
L(z k , λ∗ ) f (x ∗ ) + 12 tk2 w T ∇x2x L(x ∗ , λ∗ ) + o(tk2 ).
(12.63)
By substituting into (12.63), we obtain
f (z k ) f (x ∗ ) + 12 tk2 w T ∇x2x L(x ∗ , λ∗ )w + o(tk2 ).
(12.64)
If w T ∇x2x L(x ∗ , λ∗ )w < 0, then (12.64) would imply that f (z k ) < f (x ∗ ) for all k sufficiently
large, contradicting the fact that x ∗ is a local solution. Hence, the condition (12.57) must
hold, as claimed.
Sufficient conditions are conditions on f and ci , i ∈ E ∪ I, that ensure that x ∗ is a local
solution of the problem (12.1). (They take the opposite tack to necessary conditions, which
assume that x ∗ is a local solution and deduce properties of f and ci , for the active indices
i.) The second-order sufficient condition stated in the next theorem looks very much like
the necessary condition just discussed, but it differs in that the constraint qualification is
not required, and the inequality in (12.57) is replaced by a strict inequality.
Theorem 12.6 (Second-Order Sufficient Conditions).
Suppose that for some feasible point x ∗ ∈ IRn there is a Lagrange multiplier vector λ∗
such that the KKT conditions (12.34) are satisfied. Suppose also that
w T ∇x2x L(x ∗ , λ∗ )w > 0,
for all w ∈ C(x ∗ , λ∗ ), w 0.
(12.65)
Then x ∗ is a strict local solution for (12.1).
PROOF. First, note that the set C̄ {d ∈ C(x ∗ , λ∗ ) | d 1} is a compact subset of
C(x ∗ , λ∗ ), so by (12.65), the minimizer of d T ∇x2x L(x ∗ , λ∗ )d over this set is a strictly positive
number, say σ . Since C(x ∗ , λ∗ ) is a cone, we have that (w/w) ∈ C̄ if and only if w ∈
C(x ∗ , λ∗ ), w 0. Therefore, condition (12.65) by
w T ∇x2x L(x ∗ , λ∗ )w ≥ σ w2 ,
for all w ∈ C(x ∗ , λ∗ ),
(12.66)
for σ > 0 defined as above. (Note that this inequality holds trivially for w 0.)
We prove the result by showing that every feasible sequence {z k } approaching x ∗ has
f (z k ) ≥ f (x ∗ ) + (σ/4)z k − x ∗ 2 , for all k sufficiently large. Suppose for contradiction
that this is not the case, and that there is a sequence {z k } approaching x ∗ with
f (z k ) < f (x ∗ ) + (σ/4)z k − x ∗ 2 ,
for all k sufficiently large.
(12.67)
333
334
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
By taking a subsequence if necessary, we can identify a limiting direction d such that
zk − x ∗
d.
k→∞ z k − x ∗ lim
(12.68)
We have from Lemma 12.2(i) and Definition 12.3 that d ∈ F(x ∗ ). From (12.33) and the
facts that λi∗ ≥ 0 and ci (z k ) ≥ 0 for i ∈ I and ci (z k ) 0 for i ∈ E, we have that
L(z k , λ∗ ) f (z k ) −
λi∗ ci (z k ) ≤ f (z k ),
(12.69)
i∈A(x ∗ )
while the Taylor series approximation (12.63) from the proof of Theorem 12.5 continues to
hold.
If d were not in C(x ∗ , λ∗ ), we could identify some index j ∈ A(x ∗ ) ∩ I such that the
strict positivity condition
λ∗j ∇c j (x ∗ )T d > 0
(12.70)
is satisfied, while for the remaining indices i ∈ A(x ∗ ), we have
λi∗ ∇ci (x ∗ )T d ≥ 0.
From Taylor’s theorem and (12.68), we have for this particular value of j that
λ∗j c j (z k ) λ∗j c j (x ∗ ) + λ∗j ∇c j (x ∗ )T (z k − x ∗ ) + o(z k − x ∗ )
z k − x ∗ λ∗j ∇c j (x ∗ )T d + o(z k − x ∗ ).
Hence, from (12.69), we have that
L(z k , λ∗ ) f (z k ) −
λi∗ ci (z k )
i∈A(x ∗ )
≤ f (z k ) − λ∗j c j (z k )
≤ f (z k ) − z k − x ∗ λ∗j ∇c j (x ∗ )T d + o(z k − x ∗ ).
From the Taylor series estimate (12.63), we have meanwhile that
L(z k , λ∗ ) f (x ∗ ) + O(z k − x ∗ 2 ),
and by combining with (12.71), we obtain
f (z k ) ≥ f (x ∗ ) + z k − x ∗ λ∗j ∇c j (x ∗ )T d + o(z k − x ∗ ).
(12.71)
12.5.
SECOND-ORDER CONDITIONS
Because of (12.70), this inequality is incompatible with (12.67). We conclude that d ∈
C(x ∗ , λ∗ ), and hence d T ∇x2x L(x ∗ , λ∗ )d ≥ σ .
By combining the Taylor series estimate (12.63) with (12.69) and using (12.68), we
obtain
f (z k ) ≥ f (x ∗ ) + 12 (z k − x ∗ )T ∇x2x L(x ∗ , λ∗ )(z k − x ∗ ) + o(z k − x ∗ 2 )
f (x ∗ ) + 12 d T ∇x2x L(x ∗ , λ∗ )dz k − x ∗ 2 + o(z k − x ∗ 2 )
≥ f (x ∗ ) + (σ/2)z k − x ∗ 2 + o(z k − x ∗ 2 ).
This inequality yields the contradiction to (12.67). We conclude that every feasible sequence
{z k } approaching x ∗ must satisfy f (z k ) ≥ f (x ∗ ) + (σ/4)z k − x ∗ 2 , for all k sufficiently
large, so x ∗ is a strict local solution.
❏ EXAMPLE 12.8
(EXAMPLE 12.2, ONE MORE TIME)
We now return to Example 12.2 to check the second-order conditions for problem
(12.18). In this problem we have f (x) x1 + x2 , c1 (x) 2 − x12 − x22 , E ∅, and I {1}.
The Lagrangian is
L(x, λ) (x1 + x2 ) − λ1 (2 − x12 − x22 ),
and it is easy to show that the KKT conditions (12.34) are satisfied by x ∗ (−1, −1)T , with
λ∗1 12 . The Lagrangian Hessian at this point is
∇x2x L(x ∗ , λ∗ )
2λ∗1
0
0
2λ∗1
1
0
0
1
.
This matrix is positive definite, so it certainly satisfies the conditions of Theorem 12.6. We
conclude that x ∗ (−1, −1)T is a strict local solution for (12.18). (In fact, it is the global
solution of this problem, since, as we note later, this problem is a convex programming
problem.)
❐
❏ EXAMPLE 12.9
For a more complex example, consider the problem
min −0.1(x1 − 4)2 + x22
s.t.
x12 + x22 − 1 ≥ 0,
(12.72)
335
336
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
in which we seek to minimize a nonconvex function over the exterior of the unit circle.
Obviously, the objective function is not bounded below on the feasible region, since we can
take the feasible sequence
10
0
,
20
,
0
30
,
0
40
,
0
and note that f (x) approaches −∞ along this sequence. Therefore, no global solution
exists, but it may still be possible to identify a strict local solution on the boundary of the
constraint. We search for such a solution by using the KKT conditions (12.34) and the
second-order conditions of Theorem 12.6.
By defining the Lagrangian for (12.72) in the usual way, it is easy to verify that
∇x L(x, λ) ∇x2x L(x, λ)
−0.2(x1 − 4) − 2λ1 x1
2x2 − 2λ1 x2
,
(12.73a)
.
(12.73b)
−0.2 − 2λ1
0
0
2 − 2λ1
The point x ∗ (1, 0)T satisfies the KKT conditions with λ∗1 0.3 and the active set
A(x ∗ ) {1}. To check that the second-order sufficient conditions are satisfied at this point,
we note that
∗
∇c1 (x ) 2
,
0
so that the set C defined in (12.53) is simply
C(x ∗ , λ∗ ) {(0, w2 )T | w2 ∈ IR}.
Now, by substituting x ∗ and λ∗ into (12.73b), we have for any w ∈ C(x ∗ , λ∗ ) with w 0
that w2 0 and thus
w T ∇x2x L(x ∗ , λ∗ )w 0
w2
T −0.4
0
0
1.4
0
w2
1.4w22 > 0.
Hence, the second-order sufficient conditions are satisfied, and we conclude from
Theorem 12.6 that (1, 0)T is a strict local solution for (12.72).
❐
12.5.
SECOND-ORDER CONDITIONS
SECOND-ORDER CONDITIONS AND PROJECTED HESSIANS
The second-order conditions are sometimes stated in a form that is slightly weaker
but easier to verify than (12.57) and (12.65). This form uses a two-sided projection of the
Lagrangian Hessian ∇x2x L(x ∗ , λ∗ ) onto subspaces that are related to C(x ∗ , λ∗ ).
The simplest case is obtained when the multiplier λ∗ that satisfies the KKT conditions
(12.34) is unique (as happens, for example, when the LICQ condition holds) and strict
complementarity holds. In this case, the definition (12.53) of C(x ∗ , λ∗ ) reduces to
C(x ∗ , λ∗ ) Null ∇ci (x ∗ )T
i∈A(x ∗ )
Null A(x ∗ ),
where A(x ∗ ) is defined as in (12.37). In other words, C(x ∗ , λ∗ ) is the null space of the matrix
whose rows are the active constraint gradients at x ∗ . As in (12.39), we can define the matrix
Z with full column rank whose columns span the space C(x ∗ , λ∗ );that is,
∗
C(x ∗ , λ∗ ) {Z u | u ∈ IR|A(x )| }.
Hence, the condition (12.57) in Theorem 12.5 can be restated as
u T Z T ∇x2x L(x ∗ , λ∗ )Z u ≥ 0 for all u,
or, more succinctly,
Z T ∇x2x L(x ∗ , λ∗ )Z is positive semidefinite.
Similarly, the condition (12.65) in Theorem 12.6 can be restated as
Z T ∇x2x L(x ∗ , λ∗ )Z is positive definite.
As we show next, Z can be computed numerically, so that the positive (semi)definiteness
conditions can actually be checked by forming these matrices and finding their eigenvalues.
One way to compute the matrix Z is to apply a QR factorization to the matrix of
active constraint gradients whose null space we seek. In the simplest case above (in which
the multiplier λ∗ is unique and strictly complementary holds), we define A(x ∗ ) as in (12.37)
and write the QR factorization of its transpose as
∗ T
A(x ) Q
R
0
Q1
Q2
R
0
Q 1 R,
(12.74)
where R is a square upper triangular matrix and Q is n × n orthogonal. If R is nonsingular,
we can set Z Q 2 . If R is singular (indicating that the active constraint gradients are
linearly dependent), a slight enhancement of this procedure that makes use of column
pivoting during the QR procedure can be used to identify Z .
337
338
CHAPTER 12.
12.6
THEORY
OF
CONSTRAINED OPTIMIZATION
OTHER CONSTRAINT QUALIFICATIONS
We now reconsider constraint qualifications, the conditions discussed in Sections 12.2 and
12.4 that ensure that the linearized approximation to the feasible set captures the essential
shape of in a neighborhood of x ∗ .
One situation in which the linearized feasible direction set F(x ∗ ) is obviously an
adequate representation of the actual feasible set occurs when all the active constraints are
already linear; that is,
ci (x) aiT x + bi ,
(12.75)
for some ai ∈ IRn and bi ∈ IR. It is not difficult to prove a version of Lemma 12.2 for this
situation.
Lemma 12.7.
Suppose that at some x ∗ ∈ , all active constraints ci (·), i ∈ A(x ∗ ), are linear functions.
Then F(x ∗ ) T (x ∗ ).
PROOF. We have from Lemma 12.2 (i) that T (x ∗ ) ⊂ F(x ∗ ). To prove that F(x ∗ ) ⊂ T (x ∗ ),
we choose an arbitrary w ∈ F(x ∗ ) and show that w ∈ T (x ∗ ). By Definition 12.3 and the
form (12.75) of the constraints, we have
aiT d 0,
F(x ∗ ) d T
ai d ≥ 0,
for all i ∈ E,
for all i ∈ A(x) ∩ I
.
First, note that there is a positive scalar t¯ such that the inactive constraint remain
inactive at x ∗ + tw, for all t ∈ [0, t¯], that is,
ci (x ∗ + tw) > 0,
for all i ∈ I\A(x ∗ ) and all t ∈ [0, t¯].
Now define the sequence z k by
z k x ∗ + (t¯/k)w,
k 1, 2, . . . .
Since aiT w ≥ 0 for all i ∈ I ∩ A(x ∗ ), we have
ci (z k ) ci (z k ) − ci (x ∗ ) aiT (z k − x ∗ ) t¯ T
a w ≥ 0,
k i
for all i ∈ I ∩ A(x ∗ ),
so that z k is feasible with respect to the active inequality constraints ci , i ∈ I ∩ A(x ∗ ). By the
choice of t¯, we find that z k is also feasible with respect to the inactive inequality constraints
12.6.
OTHER CONSTRAINT QUALIFICATIONS
i ∈ I\A(x ∗ ), and it is easy to show that ci (z k ) 0 for the equality constraints i ∈ E. Hence,
z k is feasible for each k 1, 2, . . .. In addition, we have that
zk − x ∗
(t¯/k)w
w,
(t¯/k)
(t¯/k)
so that indeed w is the limiting direction of {z k }. Hence, w ∈ T (x ∗ ), and the proof is
complete.
We conclude from this result that the condition that all active constraints be linear is
another possible constraint qualification. It is neither weaker nor stronger than the LICQ
condition, that is, there are situations in which one condition is satisfied but not the other
(see Exercise 12.12).
Another useful generalization of the LICQ is the Mangasarian–Fromovitz constraint
qualification (MFCQ).
Definition 12.6 (MFCQ).
We say that the Mangasarian–Fromovitz constraint qualification (MFCQ) holds if there
exists a vector w ∈ IRn such that
∇ci (x ∗ )T w > 0,
for all i ∈ A(x ∗ ) ∩ I,
∇ci (x ∗ )T w 0,
for all i ∈ E,
and the set of equality constraint gradients {∇ci (x ∗ ), i ∈ E} is linearly independent.
Note the strict inequality involving the active inequality constraints.
The MFCQ is a weaker condition than LICQ. If LICQ is satisfied, then the system of
equalities defined by
∇ci (x ∗ )T w 1,
for all i ∈ A(x ∗ ) ∩ I,
∇ci (x ∗ )T w 0,
for all i ∈ E,
has a solution w, by full rank of the active constraint gradients. Hence, we can choose the
w of Definition 12.6 to be precisely this vector. On the other hand, it is easy to construct
examples in which the MFCQ is satisfied but the LICQ is not; see Exercise 12.13.
It is possible to prove a version of the first-order necessary condition result (Theorem 12.1) in which MFCQ replaces LICQ in the assumptions. MFCQ gives rise to the nice
property that it is equivalent to boundedness of the set of Lagrange multiplier vectors λ∗ for
which the KKT conditions (12.34) are satisfied. (In the case of LICQ, this set consists of a
unique vector λ∗ , and so is trivially bounded.)
Note that constraint qualifications are sufficient conditions for the linear approximation to be adequate, not necessary conditions. For instance, consider the set defined by x2 ≥
−x12 and x2 ≤ x12 and the feasible point x ∗ (0, 0)T . None of the constraint qualifications
339
340
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
we have discussed are satisfied, but the linear approximation F(x ∗ ) {(w1 , 0)T | w1 ∈ IR}
accurately reflects the geometry of the feasible set near x ∗ .
12.7
A GEOMETRIC VIEWPOINT
Finally, we mention an alternative first-order optimality condition that depends only on the
geometry of the feasible set and not on its particular algebraic description in terms of the
constraint functions ci , i ∈ E ∪ I. In geometric terms, our problem (12.1) can be stated as
min f (x)
subject to x ∈ ,
(12.76)
where is the feasible set.
To prove a “geometric” first-order condition, we need to define the normal cone to
the set at a feasible point x.
Definition 12.7.
The normal cone to the set at the point x ∈ is defined as
N (x) {v | v T w ≤ 0 for all w ∈ T (x)},
(12.77)
where T (x) is the tangent cone of Definition 12.2. Each vector v ∈ N (x) is said to be a
normal vector.
Geometrically, each normal vector v makes an angle of at least π/2 with every tangent
vector.
The first-order necessary condition for (12.76) is delightfully simple.
Theorem 12.8.
Suppose that x ∗ is a local minimizer of f in . Then
− ∇ f (x ∗ ) ∈ N (x ∗ ).
(12.78)
PROOF. Given any d ∈ T (x ∗ ), we have for the sequences {tk } and {z k } in Definition 12.2
that
z k ∈ ,
z k x ∗ + tk d + o(tk ),
Since x ∗ is a local solution, we must have
f (z k ) ≥ f (x ∗ )
for all k.
(12.79)
12.8.
LAGRANGE MULTIPLIERS
AND
SENSITIVITY
for all k sufficiently large. Hence, since f is continuously differentiable, we have from Taylor’s
theorem (2.4) that
f (z k ) − f (x ∗ ) tk ∇ f (x ∗ )T d + o(tk ) ≥ 0.
By dividing by tk and taking limits as k → ∞, we have
∇ f (x ∗ )T d ≥ 0.
Recall that d was an arbitrary member of T (x ∗ ), so we have −∇ f (x ∗ )T d ≤ 0 for all
d ∈ T (x ∗ ). We conclude from Definition 12.7 that −∇ f (x ∗ ) ∈ N (x ∗ ).
This result suggests a close relationship between N (x ∗ ) and the conic combination
of active constraint gradients given by (12.50). When the linear independence constraint
qualification holds, identical (to within a change of sign).
Lemma 12.9.
Suppose that the LICQ assumption (Definition 12.4) holds at x ∗ . Then t the normal cone
N (x ∗ ) is simply −N , where N is the set defined in (12.50).
PROOF. The proof follows from Farkas’ Lemma (Lemma 12.4) and Definition 12.7 of
N (x ∗ ). From Lemma 12.4, we have that
g ∈ N ⇒ g T d ≥ 0 for all d ∈ F(x ∗ ).
Since we have F(x ∗ ) T (x ∗ ) from Lemma 12.2, it follows by switching the sign of this
expression that
g ∈ −N ⇒ g T d ≤ 0 for all d ∈ T (x ∗ ).
We conclude from Definition 12.7 that N (x ∗ ) −N , as claimed.
12.8
LAGRANGE MULTIPLIERS AND SENSITIVITY
The importance of Lagrange multipliers in optimality theory should be clear, but what of
their intuitive significance? We show in this section that each Lagrange multiplier λi∗ tells us
something about the sensitivity of the optimal objective value f (x ∗ ) to the presence of the
constraint ci . To put it another way, λi∗ indicates how hard f is “pushing” or “pulling” the
solution x ∗ against the particular constraint ci .
We illustrate this point with some informal analysis. When we choose an inactive
constraint i ∈
/ A(x ∗ ) such that ci (x ∗ ) > 0, the solution x ∗ and function value f (x ∗ ) are
341
342
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
indifferent to whether this constraint is present or not. If we perturb ci by a tiny amount, it
will still be inactive and x ∗ will still be a local solution of the optimization problem. Since
λi∗ 0 from (12.34e), the Lagrange multiplier indicates accurately that constraint i is not
significant.
Suppose instead that constraint i is active, and let us perturb the right-hand-side of this
constraint a little, requiring, say, that ci (x) ≥ − ∇ci (x ∗ ) instead of ci (x) ≥ 0. Suppose
that is sufficiently small that the perturbed solution x ∗ ( ) still has the same set of active
constraints, and that the Lagrange multipliers are not much affected by the perturbation.
(These conditions can be made more rigorous with the help of strict complementarity and
second-order conditions.) We then find that
− ∇ci (x ∗ ) ci (x ∗ ( )) − ci (x ∗ ) ≈ (x ∗ ( ) − x ∗ )T ∇ci (x ∗ ),
0 c j (x ∗ ( )) − c j (x ∗ ) ≈ (x ∗ ( ) − x ∗ )T ∇c j (x ∗ ),
for all j ∈ A(x ∗ ) with j i.
The value of f (x ∗ ( )), meanwhile, can be estimated with the help of (12.34a). We have
f (x ∗ ( )) − f (x ∗ ) ≈ (x ∗ ( ) − x ∗ )T ∇ f (x ∗ )
λ∗j (x ∗ ( ) − x ∗ )T ∇c j (x ∗ )
j∈A(x ∗ )
≈ − ∇ci (x ∗ )λi∗ .
By taking limits, we see that the family of solutions x ∗ ( ) satisfies
d f (x ∗ ( ))
−λi∗ ∇ci (x ∗ ).
d
(12.80)
A sensitivity analysis of this problem would conclude that if λi∗ ∇ci (x ∗ ) is large, then the
optimal value is sensitive to the placement of the ith constraint, while if this quantity is
small, the dependence is not too strong. If λi∗ is exactly zero for some active constraint, small
perturbations to ci in some directions will hardly affect the optimal objective value at all;
the change is zero, to first order.
This discussion motivates the definition below, which classifies constraints according
to whether or not their corresponding Lagrange multiplier is zero.
Definition 12.8.
Let x ∗ be a solution of the problem (12.1), and suppose that the KKT conditions (12.34)
are satisfied. We say that an inequality constraint ci is strongly active or binding if i ∈ A(x ∗ )
and λi∗ > 0 for some Lagrange multiplier λ∗ satisfying (12.34). We say that ci is weakly active
if i ∈ A(x ∗ ) and λi∗ 0 for all λ∗ satisfying (12.34).
Note that the analysis above is independent of scaling of the individual constraints.
For instance, we might change the formulation of the problem by replacing some active
12.9.
DUALITY
constraint ci by 10ci . The new problem will actually be equivalent (that is, it has the same
feasible set and same solution), but the optimal multiplier λi∗ corresponding to ci will
be replaced by λi∗ /10. However, since ∇ci (x ∗ ) is replaced by 10∇ci (x ∗ ), the product
λi∗ ∇ci (x ∗ ) does not change. If, on the other hand, we replace the objective function f by
10 f , the multipliers λi∗ in (12.34) all will need to be replaced by 10λi∗ . Hence in (12.80) we
see that the sensitivity of f to perturbations has increased by a factor of 10, which is exactly
what we would expect.
12.9
DUALITY
In this section we present some elements of the duality theory for nonlinear programming. This theory is used to motivate and develop some important algorithms, including
the augmented Lagrangian algorithms of Chapter 17. In its full generality, duality theory
ranges beyond nonlinear programming to provide important insight into the fields of convex nonsmooth optimization and even discrete optimization. Its specialization to linear
programming proved central to the development of that area; see Chapter 13. (We note that
the discussion of linear programming duality in Section 13.1 can be read without consulting
this section first.)
Duality theory shows how we can construct an alternative problem from the functions
and data that define the original optimization problem. This alternative “dual” problem is
related to the original problem (which is sometimes referred to in this context as the “primal”
for purposes of contrast) in fascinating ways. In some cases, the dual problem is easier to
solve computationally than the original problem. In other cases, the dual can be used to
obtain easily a lower bound on the optimal value of the objective for the primal problem.
As remarked above, the dual has also been used to design algorithms for solving the primal
problem.
Our results in this section are mostly restricted to the special case of (12.1) in which
there are no equality constraints and the objective f and the negatives of the inequality
constraints −ci are all convex functions. For simplicity we assume that there are m inequality
constraints labelled 1, 2, . . . , m and rewrite (12.1) as follows:
subject to ci (x) ≥ 0, i 1, 2, . . . , m.
minn f (x)
x∈IR
If we assemble the constraints into a vector function
def
c(x) (c1 (x), c2 (x), . . . , cm (x))T ,
we can write the problem as
minn f (x)
x∈IR
subject to c(x) ≥ 0,
(12.81)
343
344
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
for which the Lagrangian function (12.16) with Lagrange multiplier vector λ ∈ IRm is
L(x, λ) f (x) − λT c(x).
We define the dual objective function q : IRn → IR as follows:
def
q(λ) inf L(x, λ).
x
(12.82)
In many problems, this infimum is −∞ for some values of λ. We define the domain of q as
the set of λ values for which q is finite, that is,
def
D {λ | q(λ) > −∞}.
(12.83)
Note that calculation of the infimum in (12.82) requires finding the global minimizer of the
function L(·, λ) for the given λ which, as we have noted in Chapter 2, may be extremely
difficult in practice. However, when f and −ci are convex functions and λ ≥ 0 (the case in
which we are most interested), the function L(·, λ) is also convex. In this situation, all local
minimizers are global minimizers (as we verify in Exercise 12.4), so computation of q(λ)
becomes a more practical proposition.
The dual problem to (12.81) is defined as follows:
maxn q(λ)
λ∈IR
subject to λ ≥ 0.
(12.84)
❏ EXAMPLE 12.10
Consider the problem
min 0.5(x12 + x22 )
(x1 ,x2 )
subject to x1 − 1 ≥ 0.
(12.85)
The Lagrangian is
L(x1 , x2 , λ1 ) 0.5(x12 + x22 ) − λ1 (x1 − 1).
If we hold λ1 fixed, this is a convex function of (x1 , x2 )T . Therefore, the infimum with
respect to (x1 , x2 )T is achieved when the partial derivatives with respect to x1 and x2 are
zero, that is,
x1 − λ1 0, x2 0.
12.9.
DUALITY
By substituting these infimal values into L(x1 , x2 , λ1 ) we obtain the dual objective (12.82):
q(λ1 ) 0.5(λ21 + 0) − λ1 (λ1 − 1) −0.5λ21 + λ1 .
Hence, the dual problem (12.84) is
max −0.5λ21 + λ1 ,
λ1 ≥0
which clearly has the solution λ1 1.
(12.86)
❐
In the remainder of this section, we show how the dual problem is related to (12.81).
Our first result concerns concavity of q.
Theorem 12.10.
The function q defined by (12.82) is concave and its domain D is convex.
PROOF. For any λ0 and λ1 in IRm , any x ∈ IRn , and any α ∈ [0, 1], we have
L(x, (1 − α)λ0 + αλ1 ) (1 − α)L(x, λ0 ) + αL(x, λ1 ).
By taking the infimum of both sides in this expression, using the definition (12.82), and
using the results that the infimum of a sum is greater than or equal to the sum of infimums,
we obtain
q((1 − α)λ0 + αλ1 ) ≥ (1 − α)q(λ0 ) + αq(λ1 ),
confirming concavity of q. If both λ0 and λ1 belong to D, this inequality implies that
q((1 − α)λ0 + αλ1 ) ≥ −∞ also, and therefore (1 − α)λ0 + αλ1 ∈ D, verifying convexity
of D.
The optimal value of the dual problem (12.84) gives a lower bound on the optimal
objective value for the primal problem (12.81). This observation is a consequence of the
following weak duality result.
Theorem 12.11 (Weak Duality).
For any x̄ feasible for (12.81) and any λ̄ ≥ 0, we have q(λ̄) ≤ f (x̄).
PROOF.
q(λ̄) inf f (x) − λ̄T c(x) ≤ f (x̄) − λ̄T c(x̄) ≤ f (x̄),
x
where the final inequality follows from λ̄ ≥ 0 and c(x̄) ≥ 0.
345
346
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
For the remaining results, we note that the KKT conditions (12.34) specialized to
(12.81) are as follows:
∇ f (x̄) − ∇c(x̄)λ̄ 0,
(12.87a)
c(x̄) ≥ 0,
(12.87b)
λ̄ ≥ 0,
(12.87c)
λ̄i ci (x̄) 0, i 1, 2, . . . , m,
(12.87d)
where ∇c(x) is the n × m matrix defined by ∇c(x) [∇c1 (x), ∇c2 (x), . . . , ∇cm (x)].
The next result shows that optimal Lagrange multipliers for (12.81) are solutions of
the dual problem (12.84) under certain conditions. It is essentially due to Wolfe [309].
Theorem 12.12.
Suppose that x̄ is a solution of (12.81) and that f and −ci , i 1, 2, . . . , m are convex
functions on IRn that are differentiable at x̄. Then any λ̄ for which (x̄, λ̄) satisfies the KKT
conditions (12.87) is a solution of (12.84).
PROOF. Suppose that (x̄, λ̄) satisfies (12.87). We have from λ̄ ≥ 0 that L(·, λ̄) is a convex
and differentiable function. Hence, for any x, we have
L(x, λ̄) ≥ L(x̄, λ̄) + ∇x L(x̄, λ̄)T (x − x̄) L(x̄, λ̄),
where the last equality follows from (12.87a). Therefore, we have
q(λ̄) inf L(x, λ̄) L(x̄, λ̄) f (x̄) − λ̄T c(x̄) f (x̄),
x
where the last equality follows from (12.87d). Since from Theorem 12.11, we have q(λ) ≤
f (x̄) for all λ ≥ 0 it follows immediately from q(λ̄) f (x̄) that λ̄ is a solution of
(12.84).
Note that if the functions are continuously differentiable and a constraint qualification
such as LICQ holds at x̄, then an optimal Lagrange multiplier is guaranteed to exist, by
Theorem 12.1.
In Example 12.10, we see that λ1 1 is both an optimal Lagrange multiplier for the
problem (12.85) and a solution of (12.86). Note too that the optimal objective for both
problems is 0.5.
We prove a partial converse of Theorem 12.12, which shows that solutions to the dual
problem (12.84) can sometimes be used to derive solutions to the original problem (12.81).
The essential condition is strict convexity of the function L(·, λ̂) for a certain value λ̂. We
note that this condition holds if either f is strictly convex (as is the case in Example 12.10)
or if ci is strictly convex for some i 1, 2, . . . , m with λ̂i > 0.
12.9.
DUALITY
Theorem 12.13.
Suppose that f and −ci , i 1, 2, . . . , m are convex and continuously differentiable on
IRn . Suppose that x̄ is a solution of (12.81) at which LICQ holds. Suppose that λ̂ solves (12.84)
and that the infimum in inf x L(x, λ̂) is attained at x̂. Assume further than L(·, λ̂) is a strictly
convex function. Then x̄ x̂ (that is, x̂ is the unique solution of (12.81)), and f (x̄) L(x̂, λ̂).
PROOF. Assume for contradiction that x̄ x̂. From Theorem 12.1, because of the LICQ
assumption, there exists λ̄ satisfying (12.87). Hence, from Theorem 12.12, we have that λ̄
also solves (12.84), so that
L(x̄, λ̄) q(λ̄) q(λ̂) L(x̂, λ̂).
Because x̂ arg minx L(x, λ̂), we have from Theorem 2.2 that ∇x L(x̂, λ̂) 0. Moreover,
by strict convexity of L(·, λ̂), it follows that
L(x̄, λ̂) − L(x̂, λ̂) > ∇x L(x̂, λ̂)T (x̄ − x̂) 0.
Hence, we have
L(x̄, λ̂) > L(x̂, λ̂) L(x̄, λ̄),
so in particular we have
−λ̂T c(x̄) > −λ̄T c(x̄) 0,
where the final equality follows from (12.87d). Since λ̂ ≥ 0 and c(x̄) ≥ 0, this yields the
contradiction, and we conclude that x̂ x̄, as claimed.
In Example 12.10, at the dual solution λ1 1, the infimum of L(x1 , x2 , λ1 ) is achieved
at (x1 , x2 ) (1, 0)T , which is the solution of the original problem (12.85).
An slightly different form of duality that is convenient for computations, known as
the Wolfe dual [309], can be stated as follows:
max L(x, λ)
(12.88a)
x,λ
subject to ∇x L(x, λ) 0,
λ ≥ 0.
(12.88b)
The following results explains the relationship of the Wolfe dual to (12.81).
Theorem 12.14.
Suppose that f and −ci , i 1, 2, . . . , m are convex and continuously differentiable on
IRn . Suppose that (x̄, λ̄) is a solution pair of (12.81) at which LICQ holds. Then (x̄, λ̄) solves
the problem (12.88).
347
348
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
PROOF. From the KKT conditions (12.87) we have that ( x̄, λ̄) satisfies (12.88b), and that
L(x̄, λ̄) f (x̄). Therefore for any pair (x, λ) that satisfies (12.88b) we have that
L(x̄, λ̄) f (x̄)
≥ f (x̄) − λT c(x̄)
L(x̄, λ)
≥ L(x, λ) + ∇x L(x, λ)T (x̄ − x)
L(x, λ),
where the second inequality follows from the convexity of L(·, λ). We have therefore shown
that (x̄, λ̄) maximizes L over the constraints (12.88b), and hence solves (12.88).
❏ EXAMPLE 12.11
(LINEAR PROGRAMMING)
An important special case of (12.81) is the linear programming problem
min c T x subject to Ax − b ≥ 0,
(12.89)
for which the dual objective is
q(λ) inf c T x − λT (Ax − b) inf (c − A T λ)T x + b T λ .
x
x
If c − A T λ 0, the infimum is clearly −∞ (we can set x to be a large negative multiple
of −(c − A T λ) to make q arbitrarily large and negative). When c − A T λ 0, on the
other hand, the dual objective is simply b T λ. In maximizing q, we can exclude λ for which
c − A T λ 0 from consideration (the maximum obviously cannot be attained at a point λ
for which q(λ) −∞). Hence, we can write the dual problem (12.84) as follows:
max b T λ
λ
subject to A T λ c, λ ≥ 0.
(12.90)
The Wolfe dual of (12.89) can be written as
max c T x − λT (Ax − b)
λ
subject to A T λ c, λ ≥ 0,
and by substituting the constraint A T λ − c 0 into the objective we obtain (12.90) again.
For some matrices A, the dual problem (12.90) may be computationally easier to solve
than the original problem (12.89). We discuss the possibilities further in Chapter 13.
❐
12.9.
❏ EXAMPLE 12.12
DUALITY
(CONVEX QUADRATIC PROGRAMMING)
Consider
min
1 T
x Gx + c T x subject to Ax − b ≥ 0,
2
(12.91)
where G is a symmetric positive definite matrix. The dual objective for this problem is
q(λ) inf L(x, λ) inf
x
x
1 T
x Gx + c T x − λT (Ax − b).
2
(12.92)
Since G is positive definite, since L(·, λ) is a strictly convex quadratic function, the infimum
is achieved when ∇x L(x, λ) 0, that is,
Gx + c − A T λ 0.
(12.93)
Hence, we can substitute for x in the infimum expression and write the dual objective
explicitly as follows:
1
q(λ) − (A T λ − c)T G −1 (A T λ − c)T + b T λ.
2
Alternatively, we can write the Wolfe dual form (12.88) by retaining x as a variable and
including the constraint (12.93) explicitly in the dual problem, to obtain
max
(λ,x)
subject to
1 T
x Gx + c T x − λT (Ax − b)
2
Gx + c − A T λ 0, λ ≥ 0.
(12.94)
To make it clearer that the objective is concave, we can use the constraint to substitute
(c − A T λ)T x −x T Gx in the objective, and rewrite the dual formulation as follows:
1
max − x T Gx + λT b, subject to Gx + c − A T λ 0, λ ≥ 0.
(λ,x)
2
(12.95)
❐
Note that the Wolfe dual form requires only positive semidefiniteness of G.
NOTES AND REFERENCES
The theory of constrained optimization is discussed in many books on numerical
optimization. The discussion in Fletcher [101, Chapter 9] is similar to ours, though a little
349
350
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
terser, and includes additional material on duality. Bertsekas [19, Chapter 3] emphasizes the
role of duality and discusses sensitivity of the solution with respect to the active constraints
in some detail. The classic treatment of Mangasarian [198] is particularly notable for its
thorough description of constraint qualifications. It also has an extensive discussion of
theorems of the alternative [198, Chapter 2], placing Farkas’ Lemma firmly in the context
of other related results.
The KKT conditions were described in a 1951 paper of Kuhn and Tucker [185],
though they were derived earlier (and independently) in an unpublished 1939 master’s
thesis of W. Karush. Lagrange multipliers and optimality conditions for general problems
(including nonsmooth problems) are described in the deep and wide-ranging article of
Rockafellar [270].
Duality theory for nonlinear programming is described in the books of Rockafellar [198] and Bertsekas [19]; the latter treatment is particularly extensive and general. The
material in Section 12.9 is adapted from these sources.
We return to our claim that the set N defined by
N {By + Ct | y ≥ 0},
(where B and C are matrices of dimension n × m and n × p, respectively, and y and t are
vectors of appropriate dimensions; see (12.45)) is a closed set. This fact is needed in the
proof of Lemma 12.4 to ensure that the solution of the projection subproblem (12.47) is
well-defined. The following technical result is well known; the proof given below is due to
R. Byrd.
Lemma 12.15.
The set N is closed.
PROOF. By splitting t into positive and negative parts, it is easy to see that
N
⎧
⎪
⎨
⎪
⎩
⎡
B
C
−C
y
⎢ +
⎣ t
t−
⎫
⎤ ⎡
⎤
y
⎪
⎬
⎥ ⎢ + ⎥
⎦ ⎣ t ⎦ ≥ 0 .
⎪
⎭
t−
Hence, we can assume without loss of generality that N has the form
N {By | y ≥ 0}.
Suppose that B has dimensions n × m.
First, we show that for any s ∈ N , we can write s B I y I with y I ≥ 0, where
I ⊂ {1, 2, . . . , m}, B I is the column submatrix of B indexed by I with full column rank,
and I has minimum cardinality. To prove this claim, we assume for contradiction that
K ⊂ {1, 2, . . . , m} is an index set with minimal cardinality such that s B K y K , y K ≥ 0, yet
12.9.
DUALITY
the columns of B K are linearly dependent. Since K is minimal, y K has no zero components.
We then have a nonzero vector w such that B K w 0. Since s B K (y K + τ w) for any τ , we
can increase or decrease τ from 0 until one or more components of y K + τ w become zero,
while the other components remain positive. We define K̄ by removing the indices from K
that correspond to zero components of y K + τ w, and define ȳ K̄ to be the vector of strictly
positive components of y K + τ w. We then have that s B K̄ ȳ K̄ and ȳ K̄ ≥ 0, contradicting
our assumption that K was the set of minimal cardinality with this property.
Now let {s k } be a sequence with s k ∈ N for all k and s k → s. We prove the lemma
by showing that s ∈ N . By the claim of the previous paragraph, for all k we can write
s k B Ik y Ikk with y Ikk ≥ 0, Ik is minimal, and the columns of B Ik are linearly independent.
Since there only finitely many possible choices of index set Ik , at least one index set occurs
infinitely often in the sequence. By choosing such an index set I , we can take a subsequence
if necessary and assume without loss of generality that Ik ≡ I for all k. We then have that
s k A I y Ik with y Ik ≥ 0 and A I has full column rank. Because of the latter property, we
have that A TI A I is invertible, so that y Ik is defined uniquely as follows:
y Ik (A TI A I )−1 A TI s k ,
k 0, 1, 2, . . . .
By taking limits and using s k → s, we have that
y Ik → y I (A TI A I )−1 A TI s,
def
and moreover y I ≥ 0, since y Ik ≥ 0 for all k. Hence we can write s B I y I with y I ≥ 0, and
therefore s ∈ N .
✐
EXERCISES
✐
12.1 The following example from [268] with a single variable x ∈ IR and a single
equality constraint shows that strict local solutions are not necessarily isolated. Consider
min x
x
2
subject to c(x) 0, where c(x) x 6 sin(1/x) 0
if x 0
0
if x 0.
(12.96)
(a) Show that the constraint function is twice continuously differentiable at all x (including
at x 0) and that the feasible points are x 0 and x 1/(kπ ) for all nonzero integers
k.
(b) Verify that each feasible point except x 0 is an isolated local solution by showing that
there is a neighborhood N around each such point within which it is the only feasible
point.
351
352
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
(c) Verify that x 0 is a global solution and a strict local solution, but not an isolated local
solution
✐
12.2 Is an isolated local solution necessarily a strict local solution? Explain.
✐
12.3 Does problem (12.4) have a finite or infinite number of local solutions? Use the
first-order optimality conditions (12.34) to justify your answer.
✐
12.4 If f is convex and the feasible region is convex, show that local solutions of
the problem (12.3) are also global solutions. Show that the set of global solutions is convex.
(Hint: See Theorem 2.5.)
✐
12.5 Let v : IRn → IRm be a smooth vector function and consider the unconstrained
optimization problems of minimizing f (x) where
f (x) v(x)∞ ,
f (x) max
i1,2,...,m
vi (x).
Reformulate these (generally nonsmooth) problems as smooth constrained optimization
problems.
✐
12.6 Can you perform a smooth reformulation as in the previous question when f
is defined by
f (x) min
i1,2,...,m
f i (x)?
(N.B. “min” not “max.”) Why or why not?
✐
12.7 Show that the vector defined by (12.15) satisfies (12.14) when the first-order
optimality condition (12.10) is not satisfied.
✐
12.8 Verify that for the sequence {z k } defined by (12.30), the function f (x) x1 +
x2 satisfies f (z k+1 ) > f (z k ) for k 2, 3, . . .. (Hint: Consider the trajectory z(s) def
(− 2 − 1/s 2 , −1/s)T and show that the function h(s) f (z(s)) has h (s) > 0 for all
s ≥ 2.)
✐
12.9 Consider the problem (12.9). Specify two feasible sequences that approach the
maximizing point (1, 1)T , and show that neither sequence is a decreasing sequence for f .
✐
12.10 Verify that neither the LICQ nor the MFCQ holds for the constraint set defined
by (12.32) at x ∗ (0, 0)T .
✐
12.11 Consider the feasible set in IR2 defined by x2 ≥ 0, x2 ≤ x12 .
(a) For x ∗ (0, 0)T , write down T (x ∗ ) and F(x ∗ ).
(b) Is LICQ satisfied at x ∗ ? Is MFCQ satisfied?
12.9.
DUALITY
(c) If the objective function is f (x) −x2 , verify that that KKT conditions (12.34) are
satisfied at x ∗ .
(d) Find a feasible sequence {z k } approaching x ∗ with f (z k ) < f (x ∗ ) for all k.
✐
12.12 It is trivial to construct an example of a feasible set and a feasible point x ∗ at
which the LICQ is satisfied but the constraints are nonlinear. Give an example of the reverse
situation, that is, where the active constraints are linear but the LICQ is not satisfied.
✐
12.13 Show that for the feasible region defined by
(x1 − 1)2 + (x2 − 1)2 ≤ 2,
(x1 − 1)2 + (x2 + 1)2 ≤ 2,
x1 ≥ 0,
the MFCQ is satisfied at x ∗ (0, 0)T but the LICQ is not satisfied.
✐
12.14 Consider the half space defined by H {x ∈ IRn | a T x + α ≥ 0} where a ∈ IRn
and α ∈ IR are given. Formulate and solve the optimization problem for finding the point
x in H that has the smallest Euclidean norm.
✐
12.15 Consider the following modification of (12.36), where t is a parameter to be
fixed prior to solving the problem:
⎡
min
x
x1 −
3
2
2
+ (x2 − t)4
s.t.
1 − x1 − x2
⎢
⎢ 1 − x1 + x2
⎢
⎢ 1+x −x
1
2
⎣
1 + x1 + x2
⎤
⎥
⎥
⎥ ≥ 0.
⎥
⎦
(12.97)
(a) For what values of t does the point x ∗ (1, 0)T satisfy the KKT conditions?
(b) Show that when t 1, only the first constraint is active at the solution, and find the
solution.
✐
12.16 (Fletcher [101]) Solve the problem
min x1 + x2 subject to x12 + x22 1
x
by eliminating the variable x2 . Show that the choice of sign for a square root operation
during the elimination process is critical; the “wrong” choice leads to an incorrect answer.
✐
12.17 Prove that when the KKT conditions (12.34) and the LICQ are satisfied at a
point x ∗ , the Lagrange multiplier λ∗ in (12.34) is unique.
353
354
CHAPTER 12.
THEORY
OF
CONSTRAINED OPTIMIZATION
✐
12.18 Consider the problem of finding the point on the parabola y 15 (x − 1)2 that
is closest to (x, y) (1, 2), in the Euclidean norm sense. We can formulate this problem as
min f (x, y) (x − 1)2 + (y − 2)2
subject to (x − 1)2 5y.
(a) Find all the KKT points for this problem. Is the LICQ satisfied?
(b) Which of these points are solutions?
(c) By directly substituting the constraint into the objective function and eliminating the
variable x, we obtain an unconstrained optimization problem. Show that the solutions
of this problem cannot be solutions of the original problem.
✐
12.19 Consider the problem
min2 f (x) −2x1 + x2
subject to
x∈IR
(1 − x1 )3 − x2
≥
0
x2 +
≥
0.
0.25x12
−1
The optimal solution is x ∗ (0, 1)T , where both constraints are active.
(a) Do the LICQ hold at this point?
(b) Are the KKT conditions satisfied?
(c) Write down the sets F(x ∗ ) and C(x ∗ , λ∗ ).
(d) Are the second-order necessary conditions satisfied? Are the second-order sufficient
conditions satisfied?
✐
12.20 Find the minima of the function f (x) x1 x2 on the unit circle x12 + x22 1.
Illustrate this problem geometrically.
✐
12.21 Find the maxima of f (x) x1 x2 over the unit disk defined by the inequality
constraint 1 − x12 − x22 ≥ 0.
✐
12.22 Show that for (12.1), the feasible set is convex if ci , i ∈ E are linear functions
and −ci , i ∈ I are convex functions.
This is page 355
Printer: Opaque this
13
CHAPTER
Linear
Programming: The
Simplex Method
Dantzig’s development of the simplex method in the late 1940s marks the start of the modern
era in optimization. This method made it possible for economists to formulate large models
and analyze them in a systematic and efficient way. Dantzig’s discovery coincided with the
development of the first electronic computers, and the simplex method became one of the
earliest important applications of this new and revolutionary technology. From those days
to the present, computer implementations of the simplex method have been continually
improved and refined. They have benefited particularly from interactions with numerical
analysis, a branch of mathematics that also came into its own with the appearance of
electronic computers, and have now reached a high level of sophistication.
356
CHAPTER 13.
THE SIMPLEX METHOD
Today, linear programming and the simplex method continue to hold sway as the most
widely used of all optimization tools. Since 1950, generations of workers in management,
economics, finance, and engineering have been trained in the techniques of formulating
linear models and solving them with simplex-based software. Often, the situations they
model are actually nonlinear, but linear programming is appealing because of the advanced
state of the software, guaranteed convergence to a global minimum, and the fact that
uncertainty in the model makes a linear model more appropriate than an overly complex
nonlinear model. Nonlinear programming may replace linear programming as the method
of choice in some applications as the nonlinear software improves, and a new class of
methods known as interior-point methods (see Chapter 14) has proved to be faster for
some linear programming problems, but the continued importance of the simplex method
is assured for the foreseeable future.
LINEAR PROGRAMMING
Linear programs have a linear objective function and linear constraints, which may
include both equalities and inequalities. The feasible set is a polytope, a convex, connected
set with flat, polygonal faces. The contours of the objective function are planar. Figure 13.1
depicts a linear program in two-dimensional space, in which the contours of the objective
function are indicated by dotted lines. The solution in this case is unique—a single vertex.
A simple reorientation of the polytope or the objective gradient c could however make the
solution non-unique; the optimal value c T x could take on the same value over an entire
edge. In higher dimensions, the set of optimal points can be a single vertex, an edge or face,
or even the entire feasible set. The problem has no solution if the feasible set is empty (the
infeasible case) or if the objective function is unbounded below on the feasible region (the
unbounded case).
Linear programs are usually stated and analyzed in the following standard form:
min c T x, subject to Ax b, x ≥ 0,
(13.1)
where c and x are vectors in IRn , b is a vector in IRm , and A is an m × n matrix. Simple
devices can be used to transform any linear program to this form. For instance, given the
problem
min c T x, subject to Ax ≤ b
(without any bounds on x), we can convert the inequality constraints to equalities by
introducing a vector of slack variables z and writing
min c T x, subject to Ax + z b, z ≥ 0.
(13.2)
This form is still not quite standard, since not all the variables are constrained to be
CHAPTER 13.
THE SIMPLEX METHOD
c
feasible polytope
optimal point x*
Figure 13.1 A linear program in two dimensions with solution at x ∗ .
nonnegative. We deal with this by splitting x into its nonnegative and nonpositive parts,
x x + − x − , where x + max(x, 0) ≥ 0 and x − max(−x, 0) ≥ 0. The problem (13.2)
can now be written as
⎡
c
⎤T ⎡
x+
⎤
⎡
⎥
⎢
⎥ ⎢
min ⎣ −c ⎦ ⎣ x − ⎦ , s.t.
0
z
A
−A
I
x+
⎤
⎡
x+
⎤
⎢ − ⎥
⎢
⎥
⎣ x ⎦ b, ⎣ x − ⎦ ≥ 0,
z
z
which clearly has the same form as (13.1).
Inequality constraints of the form x ≤ u or Ax ≥ b always can be converted to equality
constraints by adding or subtracting slack variables to make up the difference between the
left- and right-hand sides. Hence,
x ≤ u ⇔ x + w u, w ≥ 0,
Ax ≥ b ⇔ Ax − y b, y ≥ 0.
(When we subtract the variables from the left hand side, as in the second case, they are
sometimes known as surplus variables.) We can also convert a “maximize” objective max c T x
into the “minimize” form of (13.1) by simply negating c to obtain: min (−c)T x.
We say that the linear program (13.1) is infeasible if the feasible set is empty. We
say that the problem (13.1) is unbounded if the objective function is unbounded below on
the feasible region, that is, there is a sequence of points x k feasible for (13.1) such that
c T x k ↓ −∞. (Of course, unbounded problems have no solution).
357
358
CHAPTER 13.
THE SIMPLEX METHOD
Many linear programs arise from models of transshipment and distribution networks.
These problems have additional structure in their constraints; special-purpose simplex algorithms that exploit this structure are highly efficient. We do not discuss such problems further
in this book, except to note that the subject is important and complex, and that a number
of fine texts on the topic are available (see, for example, Ahuja, Magnanti, and Orlin [1]).
For the standard formulation (13.1), we will assume throughout that m < n. Otherwise, the system Ax b contains redundant rows, or is infeasible, or defines a unique point.
When m ≥ n, factorizations such as the Q R or LU factorization (see Appendix A) can be
used to transform the system Ax b to one with a coefficient matrix of full row rank.
13.1
OPTIMALITY AND DUALITY
OPTIMALITY CONDITIONS
Optimality conditions for the problem (13.1) can be derived from the theory of
Chapter 12. Only the first-order conditions—the Karush–Kuhn–Tucker (KKT) conditions—
are needed. Convexity of the problem ensures that these conditions are sufficient for a global
minimum. We do not need to refer to the second-order conditions from Chapter 12, which
are not informative in any case because the Hessian of the Lagrangian for (13.1) is zero.
The theory we developed in Chapter 12 make derivation of optimality and duality
results for linear programming much easier than in other treatments, where this theory is
developed more or less from scratch.
The KKT conditions follow from Theorem 12.1. As stated in Chapter 12, this theorem
requires linear independence of the active constraint gradients (LICQ). However, as we
noted in Section 12.6, the result continues to hold for dependent constraints provided they
are linear, as is the case here.
We partition the Lagrange multipliers for the problem (13.1) into two vectors λ and
s, where λ ∈ IRm is the multiplier vector for the equality constraints Ax b, while s ∈ IRn
is the multiplier vector for the bound constraints x ≥ 0. Using the definition (12.33), we
can write the Lagrangian function for (13.1) as
L(x, λ, s) c T x − λT (Ax − b) − s T x.
(13.3)
Applying Theorem 12.1, we find that the first-order necessary conditions for x ∗ to be a
solution of (13.1) are that there exist vectors λ and s such that
A T λ + s c,
(13.4a)
Ax b,
(13.4b)
x ≥ 0,
(13.4c)
s ≥ 0,
(13.4d)
xi si 0, i 1, 2, . . . , n.
(13.4e)
13.1.
OPTIMALITY
AND
DUALITY
The complementarity condition (13.4e), which essentially says that at least one of the
components xi and si must be zero for each i 1, 2, . . . , n, is often written in the alternative
form x T s 0. Because of the nonnegativity conditions (13.4c), (13.4d), the two forms are
identical.
Let (x ∗ , λ∗ , s ∗ ) denote a vector triple that satisfies (13.4). By combining the three
equalities (13.4a), (13.4d), and (13.4e), we find that
c T x ∗ (A T λ∗ + s ∗ )T x ∗ (Ax ∗ )T λ∗ b T λ∗ .
(13.5)
As we shall see in a moment, b T λ is the objective function for the dual problem to (13.1),
so (13.5) indicates that the primal and dual objectives are equal for vector triples (x, λ, s)
that satisfy (13.4).
It is easy to show directly that the conditions (13.4) are sufficient for x ∗ to be a global
solution of (13.1). Let x̄ be any other feasible point, so that A x̄ b and x̄ ≥ 0. Then
c T x̄ (Aλ∗ + s ∗ )T x̄ b T λ∗ + x̄ T s ∗ ≥ b T λ∗ c T x ∗ .
(13.6)
We have used (13.4) and (13.5) here; the inequality relation follows trivially from x̄ ≥ 0 and
s ∗ ≥ 0. The inequality (13.6) tells us that no other feasible point can have a lower objective
value than c T x ∗ . We can say more: The feasible point x̄ is optimal if and only if
x̄ T s ∗ 0,
since otherwise the inequality in (13.6) is strict. In other words, when si∗ > 0, then we must
have x̄i 0 for all solutions x̄ of (13.1).
THE DUAL PROBLEM
Given the data c, b, and A, which defines the problem (13.1), we can define another,
closely related, problem as follows:
max b T λ, subject to A T λ ≤ c.
(13.7)
This problem is called the dual problem for (13.1). In contrast, (13.1) is often referred to as
the primal. We can restate (13.7) in a slightly different form by introducing a vector of dual
slack variables s, and writing
max b T λ, subject to A T λ + s c, s ≥ 0.
(13.8)
The variables (λ, s) in this problem are sometimes jointly referred to collectively as dual
variables.
359
360
CHAPTER 13.
THE SIMPLEX METHOD
The primal and dual problems present two different viewpoints on the same data.
Their close relationship becomes evident when we write down the KKT conditions for
(13.7). Let us first restate (13.7) in the form
min −b T λ subject to c − A T λ ≥ 0,
to fit the formulation (12.1) from Chapter 12. By using x ∈ IRn to denote the Lagrange
multipliers for the constraints A T λ ≤ c, we see that the Lagrangian function is
L̄(λ, x) −b T λ − x T (c − A T λ).
Using Theorem 12.1 again, we find the first-order necessary conditions for λ to be optimal
for (13.7) to be that there exists x such that
Ax b,
(13.9a)
A λ ≤ c,
(13.9b)
x ≥ 0,
(13.9c)
T
xi (c − A λ)i 0, i 1, 2, . . . , n.
T
(13.9d)
Defining s c − A T λ (as in (13.8)), we find that the conditions (13.9) and (13.4) are
identical! The optimal Lagrange multipliers λ in the primal problem are the optimal variables
in the dual problem, while the optimal Lagrange multipliers x in the dual problem are the
optimal variables in the primal problem.
Analogously to (13.6), we can show that (13.9) are in fact sufficient conditions for a
solution of the dual problem (13.7). Given x ∗ and λ∗ satisfying these conditions (so that the
triple (x, λ, s) (x ∗ , λ∗ , c − A T λ∗ ) satisfies (13.4)), we have for any other dual feasible
point λ̄ (with A T λ̄ ≤ c) that
b T λ̄ (x ∗ )T A T λ̄
(x ∗ )T (A T λ̄ − c) + c T x ∗
≤ cT x ∗
because A T λ̄ − c ≤ 0 and x ∗ ≥ 0
b T λ∗
from (13.5).
Hence λ∗ achieves the maximum of the dual objective b T λ over the dual feasible region
A T λ ≤ c, so it solves the dual problem (13.7).
The primal–dual relationship is symmetric; by taking the dual of the dual problem
(13.7), we recover the primal problem (13.1). We leave the proof of this claim as an exercise.
Given a feasible vector x for the primal (satisfying Ax b and x ≥ 0) and a feasible
point (λ, s) for the dual (satisfying A T λ + s c, s ≥ 0), we have as in (13.6) that
c T x − b T λ (c − A T λ)T x s T x ≥ 0.
(13.10)
13.1.
OPTIMALITY
AND
DUALITY
Therefore we have c T x ≥ b T λ (that is, the dual objective is a lower bound on the primal
objective) when both the primal and dual variables are feasible—a result known as weak
duality.
The following strong duality result is fundamental to the theory of linear programming.
Theorem 13.1 (Strong Duality).
(i) If either problem (13.1) or (13.7) has a (finite) solution, then so does the other, and the
objective values are equal.
(ii) If either problem (13.1) or (13.7) is unbounded, then the other problem is infeasible.
PROOF. For (i), suppose that (13.1) has a finite optimal solution x ∗ . It follows from Theorem 12.1 that there are vectors λ∗ and s ∗ such that (x ∗ , λ∗ , s ∗ ) satisfies (13.4). We noted
above that (13.4) and (13.9) are equivalent, and that (13.9) are sufficient conditions for λ∗ to
be a solution of the dual problem (13.7). Moreover, it follows from (13.5) that c T x ∗ b T λ∗ ,
as claimed.
A symmetric argument holds if we start by assuming that the dual problem (13.7) has
a solution.
To prove (ii), suppose that the primal is unbounded, that is, there is a sequence of
points x k , k 1, 2, 3, . . . such that
c T x k ↓ −∞,
Ax k b,
x k ≥ 0.
Suppose too that the dual (13.7) is feasible, that is, there exists a vector λ̄ such that A T λ̄ ≤ c.
From the latter inequality together with x k ≥ 0, we have that λ̄T Ax k ≤ c T x k , and therefore
λ̄T b λ̄T Ax k ≤ c T x k ↓ −∞,
yielding a contradiction. Hence, the dual must be infeasible.
A similar argument can be used to show that unboundedness of the dual implies
infeasibility of the primal.
As we showed in the discussion following Theorem 12.1, the multiplier values λ
and s for (13.1) indicate the sensitivity of the optimal objective value to perturbations in
the constraints. In fact, the process of finding (λ, s) for a given optimal x is often called
sensitivity analysis. Considering the case of perturbations to the vector b (the right-hand side
in (13.1) and objective gradient in (13.7)), we can make an informal argument to illustrate
the sensitivity. Suppose that this small change produces small perturbations in the primal
and dual solutions, and that the vectors s and x have zeros in the same locations as s
and x, respectively. Since x and s are complementary (see (13.4e)) it follows that
0 x T s x T s (x)T s (x)T s.
361
362
CHAPTER 13.
THE SIMPLEX METHOD
We have from Theorem 13.1 that the optimal objectives of the primal and dual problems
are equal, for both the original and perturbed problems, so
c T x b T λ,
c T (x + x) (b + b)T (λ + λ).
Moreover, by feasibility of the perturbed solutions in the perturbed problems, we have
A(x + x) b + b,
A T λ −s.
Hence, the change in optimal objective due to the perturbation is as follows:
c T x (b + b)T (λ + λ) − b T λ
(b + b)T λ + (b)T λ
(x + x)T A T λ + (b)T λ
(x + x)T s + (b)T λ
(b)T λ.
In particular, if b e j , where e j is the jth unit vector in IRm , we have for all sufficiently
small that
c T x λ j .
(13.11)
That is, the change in optimal objective is λ j times the size of the perturbation to b j , if the
perturbation is small.
13.2
GEOMETRY OF THE FEASIBLE SET
BASES AND BASIC FEASIBLE POINTS
We assume for the remainder of the chapter that
The matrix A in (13.1) has full row rank.
(13.12)
In practice, a preprocessing phase is applied to the user-supplied data to remove some
redundancies from the given constraints and eliminate some of the variables. Reformulation
by adding slack, surplus, and artificial variables can also result in A satisfying the property
(13.12) .
Each iterate generated by the simplex method is a basic feasible point of (13.1). A
vector x is a basic feasible point if it is feasible and if there exists a subset B of the index set
{1, 2, . . . , n} such that
13.2.
GEOMETRY
OF THE
FEASIBLE SET
• B contains exactly m indices;
• i∈
/ B ⇒ xi 0 (that is, the bound xi ≥ 0 can be inactive only if i ∈ B);
• The m × m matrix B defined by
B [Ai ]i∈B
(13.13)
is nonsingular, where Ai is the ith column of A.
A set B satisfying these properties is called a basis for the problem (13.1). The corresponding
matrix B is called the basis matrix.
The simplex method’s strategy of examining only basic feasible points will converge
to a solution of (13.1) only if
(a) the problem has basic feasible points; and
(b) at least one such point is a basic optimal point, that is, a solution of (13.1) that is also a
basic feasible point.
Happily, both (a) and (b) are true under reasonable assumptions, as the following result
(sometimes known as the fundamental theorem of linear programming) shows.
Theorem 13.2.
(i) If (13.1) has a nonempty feasible region, then there is at least one basic feasible point;
(ii) If (13.1) has solutions, then at least one such solution is a basic optimal point.
(iii) If (13.1) is feasible and bounded, then it has an optimal solution.
PROOF. Among all feasible vectors x, choose one with the minimal number of nonzero
components, and denote this number by p. Without loss of generality, assume that the
nonzeros are x1 , x2 , . . . , x p , so we have
p
Ai xi b.
i1
Suppose first that the columns A1 , A2 , . . . , A p are linearly dependent. Then we can
express one of them ( A p , say) in terms of the others, and write
Ap p−1
i1
Ai z i ,
(13.14)
363
364
CHAPTER 13.
THE SIMPLEX METHOD
for some scalars z 1 , z 2 , . . . , z p−1 . It is easy to check that the vector
x( ) x + (z 1 , z 2 , . . . , z p−1 , −1, 0, 0, . . . , 0)T x + z
(13.15)
satisfies Ax( ) b for any scalar . In addition, since xi > 0 for i 1, 2, . . . , p, we also
have xi ( ) > 0 for the same indices i 1, 2, . . . , p and all sufficiently small in magnitude.
However, there is a value ¯ ∈ (0, x p ] such that xi (¯ ) 0 for some i 1, 2, . . . , p. Hence,
x(¯ ) is feasible and has at most p − 1 nonzero components, contradicting our choice of p
as the minimal number of nonzeros.
Therefore, columns A1 , A2 , . . . , A p must be linearly independent, and so p ≤ m. If
p m, we are done, since then x is a basic feasible point and B is simply {1, 2, . . . , m}.
Otherwise p < m and, because A has full row rank, we can choose m − p columns from
among A p+1 , A p+2 , . . . , An to build up a set of m linearly independent vectors. We construct
B by adding the corresponding indices to {1, 2, . . . , p}. The proof of (i) is complete.
The proof of (ii) is quite similar. Let x ∗ be a solution with a minimal number of
nonzero components p, and assume again that x1∗ , x2∗ , . . . , x ∗p are the nonzeros. If the
columns A1 , A2 , . . . , A p are linearly dependent, we define
x ∗ ( ) x ∗ + z,
where z is chosen exactly as in (13.14), (13.15). It is easy to check that x ∗ ( ) will be feasible
for all sufficiently small, both positive and negative. Hence, since x ∗ is optimal, we must
have
c T (x ∗ + z) ≥ c T x ∗ ⇒
cT z ≥ 0
for all sufficiently small (positive and negative). Therefore, c T z 0 and so c T x ∗ ( ) c T x ∗
for all . The same logic as in the proof of (i) can be applied to find ¯ > 0 such that x ∗ (¯ ) is
feasible and optimal, with at most p − 1 nonzero components. This contradicts our choice
of p as the minimal number of nonzeros, so the columns A1 , A2 , . . . , A p must be linearly
independent. We can now apply the same reasoning as above to conclude that x ∗ is already
a basic feasible point and therefore a basic optimal point.
The final statement (iii) is a consequence of finite termination of the simplex method.
We comment on the latter property in the next section.
The terminology we use here is not quite standard, as the following table shows:
our terminology
basic feasible point
basic optimal point
terminology used elsewhere
basic feasible solution
optimal basic feasible solution
The standard terms arose because “solution” and “feasible solution” were originally used
as synonyms for “feasible point.” However, as the discipline of optimization developed,
13.2.
*
GEOMETRY
OF THE
FEASIBLE SET
*
*
*
*
*
*
*
*
*
Figure 13.2
Vertices of a
three-dimensional polytope
(indicated by ∗).
the word “solution” took on a more specific and intuitive meaning (as in “solution to the
problem”). We maintain consistency with the rest of the book by following this more modern
usage.
VERTICES OF THE FEASIBLE POLYTOPE
The feasible set defined by the linear constraints is a polytope, and the vertices of this
polytope are the points that do not lie on a straight line between two other points in the set.
Geometrically, they are easily recognizable; see Figure 13.2. Algebraically, the vertices are
exactly the basic feasible points defined above. We therefore have an important relationship
between the algebraic and geometric viewpoints and a useful aid to understanding how the
simplex method works.
Theorem 13.3.
All basic feasible points for (13.1) are vertices of the feasible polytope {x | Ax b, x ≥ 0},
and vice versa.
PROOF. Let x be a basic feasible point and assume without loss of generality that B {1, 2, . . . , m}. The matrix B [Ai ]i1,2,...,m is therefore nonsingular, and
xm+1 xm+2 · · · xn 0.
(13.16)
Suppose that x lies on a straight line between two other feasible points y and z. Then we
can find α ∈ (0, 1) such that x αy + (1 − α)z. Because of (13.16) and the fact that α and
1 − α are both positive, we must have yi z i 0 for i m + 1, m + 2, . . . , n. Writing
xB (x1 , x2 , . . . , xm )T and defining yB and z B likewise, we have from Ax Ay Az b
365
366
CHAPTER 13.
THE SIMPLEX METHOD
that
BxB ByB Bz B b,
and so, by nonsingularity of B, we have xB yB z B . Therefore, x y z, contradicting
our assertion that y and z are two feasible points other than x. Therefore, x is a vertex.
Conversely, let x be a vertex of the feasible polytope, and suppose that the nonzero
components of x are x1 , x2 , . . . , x p . If the corresponding columns A1 , A2 , . . . , A p are
linearly dependent, then we can construct the vector x( ) x + z as in (13.15). Since
x( ) is feasible for all with sufficiently small magnitude, we can define ˆ > 0 such that
x(ˆ ) and x(−ˆ ) are both feasible. Since x x(0) obviously lies on a straight line between
these two points, it cannot be a vertex. Hence our assertion that A1 , A2 , . . . , A p are linearly
dependent must be incorrect, so these columns must be linearly independent and p ≤ m.
If p < m, and since A has full row rank, we can add m − p indices to {1, 2, . . . , p} to
form a basis B, for which x is the corresponding basic feasible point. This completes our
proof.
We conclude this discussion of the geometry of the feasible set with a definition of
degeneracy. This term has a variety of meanings in optimization, as we discuss in Chapter 16.
For the purposes of this chapter, we use the following definition.
Definition 13.1 (Degeneracy).
A basis B is said to be degenerate if xi 0 for some i ∈ B, where x is the basic feasible
solution corresponding to B. A linear program (13.1) is said to be degenerate if it has at least
one degenerate basis.
13.3
THE SIMPLEX METHOD
OUTLINE
In this section we give a detailed description of the simplex method for (13.1). There
are actually a number of variants the simplex method; the one described here is sometimes
known as the revised simplex method. (We will describe an alternative known as the dual
simplex method in Section 13.6.)
As we described above, all iterates of the simplex method are basic feasible points for
(13.1) and therefore vertices of the feasible polytope. Most steps consist of a move from one
vertex to an adjacent one for which the basis B differs in exactly one component. On most
steps (but not all), the value of the primal objective function c T x is decreased. Another type
of step occurs when the problem is unbounded: The step is an edge along which the objective
function is reduced, and along which we can move infinitely far without ever reaching a
vertex.
13.3.
THE SIMPLEX METHOD
The major issue at each simplex iteration is to decide which index to remove from the
basis B. Unless the step is a direction of unboundedness, a single index must be removed
from B and replaced by another from outside B. We can gain some insight into how this
decision is made by looking again at the KKT conditions (13.4).
From B and (13.4), we can derive values for not just the primal variable x but also
the dual variables (λ, s), as we now show. First, define the nonbasic index set N as the
complement of B, that is,
N {1, 2, . . . , n}\B.
(13.17)
Just as B is the basic matrix, whose columns are Ai for i ∈ B, we use N to denote the nonbasic
matrix N [Ai ]i∈N . We also partition the n-element vectors x, s, and c according to the
index sets B and N , using the notation
xB [xi ]i∈B ,
xN [xi ]i∈N ,
sB [si ]i∈B ,
sN [si ]i∈N ,
cB [ci ]i∈B ,
cN [ci ]i∈N .
From the KKT condition (13.4b), we have that
Ax BxB + N xN b.
The primal variable x for this simplex iterate is defined as
xB B −1 b,
xN 0.
(13.18)
Since we are dealing only with basic feasible points, we know that B is nonsingular and
that xB ≥ 0, so this choice of x satisfies two of the KKT conditions: the equality constraints
(13.4b) and the nonnegativity condition (13.4c).
We choose s to satisfy the complementarity condition (13.4e) by setting sB 0. The
remaining components λ and sN can be found by partitioning this condition into cB and cN
components and using sB 0 to obtain
B T λ cB ,
N T λ + sN cN .
(13.19)
Since B is square and nonsingular, the first equation uniquely defines λ as
λ B −T cB .
(13.20)
The second equation in (13.19) implies a value for sN :
sN cN − N T λ cN − (B −1 N )T cB .
(13.21)
367
368
CHAPTER 13.
THE SIMPLEX METHOD
Computation of the vector sN is often referred to as pricing. The components of sN are often
called the reduced costs of the nonbasic variables xN .
The only KKT condition that we have not enforced explicitly is the nonnegativity
condition s ≥ 0. The basic components sB certainly satisfy this condition, by our choice
sB 0. If the vector sN defined by (13.21) also satisfies sN ≥ 0, we have found an optimal
vector triple (x, λ, s), so the algorithm can terminate and declare success. Usually, however,
one or more of the components of sN are negative. The new index to enter the basis B—the
entering index—is chosen to be one of the indices q ∈ N for which sq < 0. As we show below,
the objective c T x will decrease when we allow xq to become positive if and only if (i) sq < 0
and (ii) it is possible to increase xq away from zero while maintaining feasibility of x. Our
procedure for altering B and changing x and s can be described accordingly as follows:
• allow xq to increase from zero during the next step;
• fix all other components of xN at zero, and figure out the effect of increasing xq on the
current basic vector xB , given that we want to stay feasible with respect to the equality
constraints Ax b;
• keep increasing xq until one of the components of xB (x p , say) is driven to zero, or
determining that no such component exists (the unbounded case);
• remove index p (known as the leaving index) from B and replace it with the entering
index q.
This process of selecting entering and leaving indices, and performing the algebraic
operations necessary to keep track of the values of the variables x, λ, and s, is sometimes
known as pivoting.
We now formalize the pivoting procedure in algebraic terms. Since both the new
iterate x + and the current iterate x should satisfy Ax b, and since xN 0 and xi+ 0 for
i ∈ N \{q}, we have
Ax + BxB+ + Aq xq+ BxB Ax.
By multiplying this expression by B −1 and rearranging, we obtain
xB+ xB − B −1 Aq xq+ .
(13.22)
Geometrically speaking, (13.22) is usually a move along an edge of the feasible polytope that
decreases c T x. We continue to move along this edge until a new vertex is encountered. At
this vertex, a new constraint x p ≥ 0 must have become active, that is, one of the components
x p , p ∈ B, has decreased to zero. We then remove this index p from the basis B and replace
it by q.
13.3.
THE SIMPLEX METHOD
We now show how the step defined by (13.22) affects the value of c T x. From (13.22),
we have
c T x + cBT xB+ + cq xq+ cBT xB − cBT B −1 Aq xq+ + cq xq+ .
(13.23)
From (13.20) we have cBT B −1 λT , while from the second equation in (13.19), since q ∈ N ,
we have AqT λ cq − sq . Therefore,
cBT B −1 Aq xq+ λT Aq xq+ (cq − sq )xq+ ,
so by substituting in (13.23) we obtain
c T x + cBT xB − (cq − sq )xq+ + cq xq+ c T x + sq xq+ .
(13.24)
Since q was chosen to have sq < 0, it follows that the step (13.22) produces a decrease in
the primal objective function c T x whenever xq+ > 0.
It is possible that we can increase xq+ to ∞ without ever encountering a new vertex.
In other words, the constraint xB+ xB − B −1 Aq xq+ ≥ 0 holds for all positive values of xq+ .
When this happens, the linear program is unbounded; the simplex method has identified a
ray that lies entirely within the feasible polytope along which the objective c T x decreases
to −∞.
Figure 13.3 shows a path traversed by the simplex method for a problem in IR2 . In this
example, the optimal vertex x ∗ is found in three steps.
If the basis B is nondegenerate (see Definition 13.1), then we are guaranteed that
xq+ > 0, so we can be assured of a strict decrease in the objective function c T x at this step. If
c
0
1
simplex path
2
3
Figure 13.3
Simplex iterates for
a two-dimensional
problem.
369
370
CHAPTER 13.
THE SIMPLEX METHOD
the problem (13.1) is nondegenerate, we can ensure a decrease in c T x at every step, and can
therefore prove the following result concerning termination of the simplex method.
Theorem 13.4.
Provided that the linear program (13.1) is nondegenerate and bounded, the simplex
method terminates at a basic optimal point.
PROOF. The simplex method cannot visit the same basic feasible point x at two different
iterations, because it attains a strict decrease at each iteration. Since the number of possible
bases B is finite (there are only a finite number of ways to choose a subset of m indices
from {1, 2, . . . , n}), and since each basis defines a single basic feasible point, there are only
a finite number of basic feasible points. Hence, the number of iterations is finite. Moreover,
since the method is always able to take a step away from a nonoptimal basic feasible point,
and since the problem is not unbounded, the method must terminate at a basic optimal
point.
This result gives us a proof of Theorem 13.2 (iii) in the case in which the linear
program is nondegenerate. The proof of finite termination is considerably more complex
when nondegeneracy of (13.1) is not assumed, as we discuss at the end of Section 13.5.
A SINGLE STEP OF THE METHOD
We have covered most of the mechanics of taking a single step of the simplex method.
To make subsequent discussions easier to follow, we summarize our description.
Procedure 13.1 (One Step of Simplex).
Given B, N , xB B −1 b ≥ 0, xN 0;
Solve B T λ cB for λ,
Compute sN cN − N T λ; (∗ pricing ∗ )
if sN ≥ 0
stop; (∗ optimal point found ∗ )
Select q ∈ N with sq < 0 as the entering index;
Solve Bd Aq for d;
if d ≤ 0
stop; (∗ problem is unbounded ∗ )
Calculate xq+ mini | di >0 (xB )i /di , and use p to denote the minimizing i;
Update xB+ xB − d xq+ , xN+ (0, . . . , 0, xq+ , 0, . . . , 0)T ;
Change B by adding q and removing the basic variable corresponding to column p of B.
We illustrate this procedure with a simple example.
13.3.
THE SIMPLEX METHOD
❏ EXAMPLE 13.1
Consider the problem
min −4x1 − 2x2
subject to
x1 + x2 + x3 5,
2x1 + (1/2)x2 + x4 8,
x ≥ 0.
Suppose we start with the basis B {3, 4}, for which we have
xB x3
x4
5
,
8
λ
0
0
sN ,
s1
s2
−3
,
−2
and an objective value of c T x 0. Since both elements of sN are negative, we could choose
either 1 or 2 to be the entering variable. Suppose we choose q 1. We obtain d (1, 2)T ,
so we cannot (yet) conclude that the problem is unbounded. By performing the ratio
calculation, we find that p 2 (corresponding to the index 4) and x1+ 4. We update the
basic and nonbasic index sets to B {3, 1} and N {4, 2}, and move to the next iteration.
At the second iteration, we have
xB x3
x1
1
4
,
λ
0
−3/2
sN ,
s4
s2
3/2
−5/4
,
with an objective value of −12. We see that sN has one negative component, corresponding
to the index q 2, so we select this index to enter the basis. We obtain d (3/2, −1/2)T ,
so again we do not detect unboundedness. Continuing, we find that the maximum value of
x2+ is 4/3, and that p 1, which indicates that index 3 will leave the basis B. We update the
index sets to B {2, 1} and N {4, 3} and continue.
At the start of the third iteration, we have
xB x2
x1
4/3
11/3
,
λ
−5/3
−2/3
,
sN s4
s3
7/3
5/3
,
with an objective value of c T x −41/3. We see that sN ≥ 0, so the optimality test is
satisfied, and we terminate.
❐
We need to flesh out Procedure 13.1 with specifics of three important aspects of the
implementation:
371
372
CHAPTER 13.
THE SIMPLEX METHOD
• Linear algebra issues—maintaining an LU factorization of B that can be used to solve
for λ and d.
• Selection of the entering index q from among the negative components of sN . (In
general, there are many such components.)
• Handling of degenerate bases and degenerate steps, in which it is not possible to
choose a positive value of xq+ without violating feasibility.
Proper handling of these issues is crucial to the efficiency of a simplex implementation. We
give some details in the next three sections.
13.4
LINEAR ALGEBRA IN THE SIMPLEX METHOD
We have to solve two linear systems involving the matrix B at each step; namely,
B T λ cB ,
Bd Aq .
(13.25)
We never calculate the inverse basis matrix B −1 explicitly just to solve these systems. Instead,
we calculate or maintain some factorization of B—usually an LU factorization—and use
triangular substitutions with the factors to recover λ and d. It is less expensive to update the
factorization than to calculate it afresh at each iteration because the basis matrix B changes
by just a single column between iterations.
The standard factorization/updating procedures start with an LU factorization of B
at the first iteration of the simplex algorithm. Since in practical applications B is large
and sparse, its rows and columns are rearranged during the factorization to maintain both
numerical stability and sparsity of the L and U factors. One successful pivot strategy that
trades off between these two aims was proposed by Markowitz in 1957 [202]; it is still used
as the basis of many practical sparse LU algorithms. Other considerations may also enter
into our choice of row and column reordering of B. For example, it may help to improve the
efficiency of the updating procedure if as many as possible of the leading columns of U contain just a single nonzero, on the diagonal. Many heuristics have been devised for choosing
row and column permutations that produce this and other desirable structural features.
Let us assume for simplicity that row and column permutations are already
incorporated in B, so that we write the initial LU factorization as
LU B,
(13.26)
(L is unit lower triangular, U is upper triangular). The system Bd Aq can then be solved
by the following two-step procedure:
L d̄ Aq ,
U d d̄.
(13.27)
13.4.
LINEAR ALGEBRA
IN THE
SIMPLEX METHOD
column p
Figure 13.4 Left: L −1 B + , which is upper triangular except for the column occupied
by A p . Right: After cyclic row and column permutation P1 , the non–upper triangular
part of P1 L −1 B + P1T appears in the last row.
Similarly, the system B T λ cB is solved by performing the following two triangular
substitutions:
U T λ̄ cB ,
L T λ λ̄.
We now discuss a procedure for updating the factors L and U after one step of the
simplex method, when the index p is removed from the basis B and replaced by the index
q. The corresponding change to the basis matrix B is that the column B p is removed from
B and replaced by Aq . We call the resulting matrix B + and note that if we rewrite (13.26) as
U L −1 B, the modified matrix L −1 B + will be upper triangular except in column p. That
is, L −1 B + has the form shown on the left in Figure 13.4.
We now perform a cyclic permutation that moves column p to the last column position
m and moves columns p + 1, p + 2, . . . , m one position to the left to make room for it. If we
apply the same permutation to rows p through m, the net effect is to move the non-upper
triangular part to the last row of the matrix, as shown in Figure 13.4. If we denote the
permutation matrix by P1 , the matrix illustrated at right in Figure 13.4 is P1 L −1 B + P1T .
Finally, we perform sparse Gaussian elimination on the matrix P1 L −1 B + P1T to restore
upper triangular form. That is, we find L 1 and U1 (lower and upper triangular, respectively)
such that
P1 L −1 B + P1T L 1 U1 .
(13.28)
It is easy to show that L 1 and U1 have a simple form. The lower triangular matrix L 1
differs from the identity only in the last row, while U1 is identical to P1 L −1 B + P1T except
that the (m, m) element is changed and the off-diagonal elements in the last row are
eliminated.
373
374
CHAPTER 13.
THE SIMPLEX METHOD
We give details of this process for the case of m 5. Using the notation
⎡
⎢
⎢
⎢
⎢
−1
L B U ⎢
⎢
⎢
⎣
u 11
u 12
u 13
u 14
u 22
u 23
u 24
u 33
u 34
u 15
⎤
⎡
⎥
u 25 ⎥
⎥
⎥
u 35 ⎥ ,
⎥
u 45 ⎥
⎦
u 55
u 44
w1
⎤
⎥
⎢
⎢ w2 ⎥
⎥
⎢
⎥
⎢
L −1 Aq ⎢ w3 ⎥ ,
⎥
⎢
⎢ w ⎥
⎣ 4 ⎦
w5
and supposing that p 2 (so that the second column is replaced by L −1 Aq ), we have
⎡
u 11
⎢
⎢
⎢
⎢
−1 +
L B ⎢
⎢
⎢
⎣
w1
u 13
u 14
w2
u 23
u 24
w3
u 33
u 34
w4
u 44
w5
u 15
⎤
⎥
u 25 ⎥
⎥
⎥
u 35 ⎥ .
⎥
u 45 ⎥
⎦
u 55
After the cyclic permutation P1 , we have
⎡
P1 L −1 B + P1 T
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
u 11
u 13
u 14
u 15
u 33
u 34
u 35
u 44
u 45
u 55
u 23
u 24
u 25
w1
⎤
⎥
w3 ⎥
⎥
⎥
w4 ⎥ .
⎥
w5 ⎥
⎦
w2
(13.29)
The factors L 1 and U1 are now as follows:
⎡
⎢
⎢
⎢
⎢
L1 ⎢
⎢
⎢
⎣
⎤
1
⎥
⎥
⎥
⎥
⎥,
⎥
⎥
⎦
1
1
1
0
l52
l53
l54
⎡
⎢
⎢
⎢
⎢
U1 ⎢
⎢
⎢
⎣
u 11
u 13
u 14
u 15
u 33
u 34
u 35
u 44
u 45
u 55
1
w1
⎤
⎥
w3 ⎥
⎥
⎥
w4 ⎥ ,
⎥
w5 ⎥
⎦
ŵ2
(13.30)
for certain values of l52 , l53 , l54 , and ŵ2 (see Exercise 13.10).
The result of this updating process is the factorization (13.28), which we can rewrite
as follows:
B + L +U + ,
where L + L P1T L 1 , U + U1 P1 .
(13.31)
13.5.
OTHER IMPORTANT DETAILS
There is no need to calculate L + and U + explicitly. Rather, the nonzero elements in L 1 and
the last column of U1 , and the permutation information in P1 , can be stored in compact
form, so that triangular substitutions involving L + and U + can be performed by applying
a number of permutations and sparse triangular substitutions involving these factors. The
factorization updates from subsequent simplex steps are stored and applied in a similar
fashion.
The procedure we have just outlined is due to Forrest and Tomlin [110]. It is quite
efficient, because it requires the storage of little data at each update and does not require much
movement of data in memory. Its major disadvantage is possible numerical instability. Large
elements in the factors of a matrix are a sure indicator of instability, and the multipliers in
the L 1 factor (l52 in (13.30), for example) may be very large. An earlier scheme of Bartels and
Golub [12] allowed swapping of rows to avoid these problems. For instance, if |u 33 | < |u 23 |
in (13.29), we could swap rows 2 and 5 to ensure that the subsequent multiplier l52 in the L 1
factor does not exceed 1 in magnitude. This improved stability comes at a price: The lower
right corner of the upper triangular factor may become more dense during each update.
Although the update information for each iteration (the permutation matrices and
the sparse triangular factors) can often be stored in a highly compact form, the total amount
of space may build up to unreasonable levels after many such updates have been performed.
As the number of updates builds up, so does the time needed to solve for the vectors
d and λ in Procedure 13.1. If an unstable updating procedure is used, numerical errors
may also come into play, blocking further progress by the simplex algorithm. For all these
reasons, most simplex implementations periodically calculate a fresh LU factorization of
the current basis matrix B and discard the accumulated updates. The new factorization
uses the same permutation strategies that we apply to the very first factorization, which
balance the requirements of stability, sparsity, and structure.
13.5
OTHER IMPORTANT DETAILS
PRICING AND SELECTION OF THE ENTERING INDEX
There are usually many negative components of sN at each step. How do we choose
one of these to become the index that enters the basis? Ideally, we would like to choose the
sequence of entering indices q that gets us to the solution x ∗ in the fewest possible steps,
but we rarely have the global perspective needed to implement this strategy. Instead, we use
more shortsighted but practical strategies that obtain a significant decrease in c T x on just
the present iteration. There is usually a tradeoff between the effort spent on finding a good
entering index and the amount of decrease in c T x resulting from this choice. Different pivot
strategies resolve this tradeoff in different ways.
Dantzig’s original selection rule is one of the simplest. It chooses q such that sq is
the most negative component of sN N T λ. This rule, which is motivated by (13.24), gives
the maximum improvement in c T x per unit increase in the entering variable xq . A large
375
376
CHAPTER 13.
THE SIMPLEX METHOD
reduction in c T x is not guaranteed, however. It could be that we can increase xq+ only a tiny
amount from zero (or not at all) before reaching the next vertex.
Calculation of the entire vector sN from (13.21) requires a multiplication by N T , which
can be expensive when the matrix N is very large. Partial pricing strategies calculate only a
subvector of sN and make the choice of entering variable from among the negative entries in
this subvector. To give all the indices in N a chance to enter the basis, these strategies cycle
through the nonbasic elements, periodically changing the subvector of sN they evaluate so
that no nonbasic index is ignored for too long.
Neither of these strategies guarantees that we can make a substantial move along the
chosen edge before reaching a new vertex. Multiple pricing strategies are more thorough: For
a small subset of indices q ∈ N , they evaluate sq and, if sq < 0, the maximum value of xq+
that maintains feasibility of x + and the consequent change sq xq+ in the objective function
(see (13.24)). Calculation of xq+ requires evaluation of d B −1 Aq as in Procedure 13.1,
which is not cheap. Subsequent iterations deal with this same index subset until we reach
an iteration at which all sq are nonnegative for q in the subset. At this point, the full vector
sN is computed, a new subset of nonbasic indices is chosen, and the cycle begins again. This
approach has the advantage that the columns of the matrix N outside the current subset of
priced components need not be accessed at all, so memory access in the implementation is
quite localized.
Naturally, it is possible to devise heuristics that combine partial and multiple pricing
in various imaginative ways.
A sophisticated rule known as steepest edge chooses the “most downhill” direction
from among all the candidates—the one that produces the largest decrease in c T x per unit
distance moved along the edge. (By contrast, Dantzig’s rule maximizes the decrease in c T x
per unit change in xq+ , which is not the same thing, as a small change in xq+ can correspond
to a large distance moved along the edge.) During the pivoting step, the overall change in x is
+
x xB+
xN+
xB
xN
+
−B −1 Aq
eq
xq+ x + ηq xq+ ,
(13.32)
where eq is the unit vector with a 1 in the position corresponding to the index q ∈ N and
zeros elsewhere, and the vector ηq is defined as
ηq −B −1 Aq
eq
−d
eq
;
(13.33)
see (13.25). The change in c T x per unit step along ηq is given by
c T ηq
.
ηq The steepest-edge rule chooses q ∈ N to minimize this quantity.
(13.34)
13.5.
OTHER IMPORTANT DETAILS
If we had to compute each ηi by solving Bd Ai for each i ∈ N , the steepest-edge
strategy would be prohibitively expensive. Goldfarb and Reid [134] showed that the measure
(13.34) of edge steepness for all indices i ∈ N can, in fact, be updated quite economically
at each iteration. We outline their steepest-edge procedure by showing how each c T ηi and
ηi can be updated at the current iteration.
First, note that we already know the numerator c T ηi in (13.34) without calculating
ηi , because by taking the inner product of (13.32) with c and using (13.24), we have that
c T ηi si . To investigate the change in denominator ηi at this step, we define γi ηi 2 ,
where this quantity is defined before and after the update as follows:
γi ηi 2 B −1 Ai 2 + 1,
γi+
ηi+ 2
+ −1
(B )
(13.35a)
Ai + 1.
2
(13.35b)
Assume without loss of generality that the entering column Aq replaces the first column of
the basis matrix B (that is, p 1), and that this column corresponds to the index t. We can
then express the update to B as follows:
B + B + (Aq − At )e1T B + (Aq − Be1 )e1T ,
(13.36)
where e1 (1, 0, 0, . . . , 0)T . By applying the Sherman–Morrison formula (A.27) to the
rank-one update formula in (13.36), we obtain
(B + )−1 B −1 −
(B −1 Aq − e1 )e1T B −1
(d − e1 )e1T B −1
−1
B
,
−
1 + e1T (B −1 Aq − e1 )
e1T d
where again we have used the fact that d B −1 Aq (see (13.25)). Therefore, we have that
(B + )−1 Ai B −1 Ai −
e1T B −1 Ai
(d − e1 ).
e1T d
By substituting for (B + )−1 Ai in (13.35) and performing some simple manipulation, we
obtain
γi+ γi − 2
e1T B −1 Ai
e1T d
AiT B −T d +
e1T B −1 Ai
e1T d
2
γq .
(13.37)
Once we solve the following two linear systems to obtain d̂ and r :
B T d̂ d,
B T r e1 .
(13.38)
377
378
CHAPTER 13.
THE SIMPLEX METHOD
The formula (13.37) then becomes
γi+ γi − 2
r T Ai
r T Aq
d̂ T Ai +
r T Ai
r T Aq
2
γq .
(13.39)
Hence, the entire set of γi values, for i ∈ N with i q, can be calculated by solving the two
systems (13.38) and then evaluating the inner products r T Ai and d̂ T Ai , for each i.
The steepest-edge strategy does not guarantee that we can take a long step before
reaching another vertex, but it has proved to be highly effective in practice.
STARTING THE SIMPLEX METHOD
The simplex method requires a basic feasible starting point x and a corresponding
initial basis B ⊂ {1, 2, . . . , n} with |B| m such that the basis matrix B defined by (13.13)
is nonsingular and xB B −1 b ≥ 0 and xN 0. The problem of finding this initial point and
basis may itself be nontrivial—in fact, its difficulty is equivalent to that of actually solving
a linear program. We describe here the two-phase approach that is commonly used to deal
with this difficulty in practical implementations.
In Phase I of this approach we set up an auxiliary linear program based on the data
of (13.1), and solve it with the simplex method. The Phase-I problem is designed so that an
initial basis and initial basic feasible point is trivial to find, and so that its solution gives a
basic feasible initial point for the second phase. In Phase II, a second linear program similar
to the original problem (13.1) is solved, with the Phase-I solution as a starting point. The
solution of the original problem (13.1) can be extracted easily from the solution of the
Phase-II problem.
In Phase I we introduce artificial variables z into (13.1) and redefine the objective
function to be the sum of these artificial variables, as follows:
min e T z, subject to Ax + E z b, (x, z) ≥ 0,
(13.40)
where z ∈ IRm , e (1, 1, . . . , 1)T , and E is a diagonal matrix whose diagonal elements are
E j j +1 if b j ≥ 0,
E j j −1 if b j < 0.
It is easy to see that the point (x, z) defined by
x 0,
z j |b j |,
j 1, 2, . . . , m,
(13.41)
is a basic feasible point for (13.40). Obviously, this point satisfies the constraints in (13.40),
while the initial basis matrix B is simply the diagonal matrix E, which is clearly nonsingular.
At any feasible point for (13.40), the artificial variables z represent the amounts by
which the constraints Ax b are violated by the x component. The objective function is
13.5.
OTHER IMPORTANT DETAILS
simply the sum of these violations, so by minimizing this sum we are forcing x to become
feasible for the original problem (13.1). It is not difficult to see that the Phase-I problem
(13.40) has an optimal objective value of zero if and only if the original problem (13.1) is
feasible, by using the following argument: If there exists a vector (x̃, z̃) that is feasible for
(13.40) such that e T z̃ 0, we must have z̃ 0, and therefore A x̃ b and x̃ ≥ 0, so x̃ is
feasible for (13.1). Conversely, if x̃ is feasible for (13.1), then the point (x̃, 0) is feasible for
(13.40) with an objective value of 0. Since the objective in (13.40) is obviously nonnegative
at all feasible points, then (x̃, 0) must be optimal for (13.40), verifying our claim.
In Phase I, we apply the simplex method to (13.40) from the initial point (13.41). This
linear program cannot be unbounded, because its objective function is bounded below by 0,
so the simplex method will terminate at an optimal point (assuming that it does not cycle; see
below). If the objective e T z is positive at this solution, we conclude by the argument above
that the original problem (13.1) is infeasible. Otherwise, the simplex method identifies a
point (x̃, z̃) with e T z̃ 0, which is also a basic feasible point for the following Phase-II
problem:
min c T x subject to Ax + z b, x ≥ 0, 0 ≥ z ≥ 0.
(13.42)
Note that this problem differs from (13.40) in that the objective function is replaced by the
original objective c T x, while upper bounds of 0 have been imposed on z. In fact, (13.42) is
equivalent to (13.1), because any solution (and indeed any feasible point) must have z 0.
We need to retain the artificial variables z in Phase II, however, since some components of z
may still be present in the optimal basis from Phase I that we are using as the initial basis for
(13.42), though of course the values z̃ j of these components must be zero. In fact, we can
modify (13.42) to include only those components of z that are present in the optimal basis
for (13.40).
The problem (13.42) is not quite in standard form because of the two-sided bounds
on z. However, it is easy to modify the simplex method described above to handle upper
and lower bounds on the variables (we omit the details). We can customize the simplex
algorithm slightly by deleting each component of z from the problem (13.42) as soon as it
is swapped out of the basis. This strategy ensures that components of z do not repeatedly
enter and leave the basis, thereby avoiding unnecessary simplex iterations.
If (x ∗ , z ∗ ) is a basic solution of (13.42), it must have z ∗ 0, and so x ∗ is a solution
of (13.1). In fact, x ∗ is a basic feasible point for (13.1), though this claim is not completely
obvious because the final basis B for the Phase-II problem may still contain components of
z ∗ , making it unsuitable as an optimal basis for (13.1). Since A has full row rank, however,
we can construct an optimal basis for (13.1) in a postprocessing phase: Extract from B any
components of z that are present, and replace them with nonbasic components of x in a
way that maintains nonsingularity of the submatrix B defined by (13.13).
A final point to note is that in many problems we do not need to add a complete set of
m artificial variables to form the Phase-I problem. This observation is particularly relevant
when slack and surplus variables have already been added to the problem formulation, as
379
380
CHAPTER 13.
THE SIMPLEX METHOD
in (13.2), to obtain a linear program with inequality constraints in standard form (13.1).
Some of these slack/surplus variables can play the roles of artificial variables, making it
unnecessary to include such variables explicitly.
We illustrate this point with the following example.
❏ EXAMPLE 13.2
Consider the inequality-constrained linear program defined by
min 3x1 + x2 + x3
subject to
2x1 + x2 + x3 ≤ 2,
x1 − x2 − x3 ≤ −1,
x ≥ 0.
By adding slack variables to both inequality constraints, we obtain the following equivalent
problem in standard form:
min 3x1 + x2 + x3
subject to
2x1 + x2 + x3 + x4 2,
x1 − x2 − x3 + x5 −1,
x ≥ 0.
By inspection, it is easy to see that the vector x (0, 0, 0, 2, 0) is feasible with respect to
the first linear constraint and the lower bound x ≥ 0, though it does not satisfy the second
constraint. Hence, in forming the Phase-I problem, we add just a single artificial variable z 2
to the second constraint and obtain
min z 2
subject to
2x1 + x2 + x3 + x4 2,
x1 − x2 − x3 + x5 − z 2 −1,
(x, z 2 ) ≥ 0.
(13.43)
(13.44)
(13.45)
(13.46)
It is easy to see that the vector (x, z 2 ) ((0, 0, 0, 2, 0), 1) is feasible with respect to (13.43).
In fact, it is a basic feasible point, since the corresponding basis matrix B is
1
0
B
,
0 −1
which is clearly nonsingular. In this example, the variable x4 plays the role of artificial variable
for the first constraint. There was no need to add an explicit artificial variable z 1 .
❐
13.5.
OTHER IMPORTANT DETAILS
DEGENERATE STEPS AND CYCLING
As noted above, the simplex method may encounter situations in which for the
entering index q, we cannot set xq+ any greater than zero in (13.22) without violating the
nonnegativity condition x + ≥ 0. By referring to Procedure 13.1, we see that these situations
arise when there is i with (xB )i 0 and di < 0, where d is defined by (13.25). Steps of
this type are called degenerate steps. On such steps, the components of x do not change
and, therefore, the objective function c T x does not decrease. However, the steps may still
be useful because they change the basis B (by replacing one index), and the updated B
may be closer to the optimal basis. In other words, the degenerate step may be laying the
groundwork for reductions in c T x on later steps.
Sometimes, however, a phenomenon known as cycling can occur. After a number of
successive degenerate steps, we may return to an earlier basis B. If we continue to apply the
algorithm from this point using the same rules for selecting entering and leaving indices,
we will repeat the same cycle ad infinitum, never converging.
Cycling was once thought to be a rare phenomenon, but in recent times it has been
observed frequently in the large linear programs that arise as relaxations of integer programming problems. Since integer programs are an important source of linear programs,
practical simplex codes usually incorporate a cycling avoidance strategy.
In the remainder of this section, we describe a perturbation strategy and its close
relative, the lexicographic strategy.
Suppose that a degenerate basis is encountered at some simplex iteration, at which
the basis is B̂ and the basis matrix is B̂, say. We consider a modified linear program in
which we add a small perturbation to the right-hand side of the constraints in (13.1), as
follows:
⎤
⎡
⎢
⎢
⎢
b( ) b + B̂ ⎢
⎢
⎣
2
..
.
⎥
⎥
⎥
⎥,
⎥
⎦
m
where is a very small positive number. This perturbation in b induces a perturbation in
the components of the basic solution vector; we have
⎤
⎡
⎢
⎢
⎢
x B̂ ( ) x B̂ + ⎢
⎢
⎣
2
..
.
m
⎥
⎥
⎥
⎥.
⎥
⎦
(13.47)
381
382
CHAPTER 13.
THE SIMPLEX METHOD
Retaining the perturbation for subsequent iterations, we see that subsequent basic
solutions have the form
⎤
⎡
xB ( ) xB + B
−1
⎢
⎢
⎢
B̂ ⎢
⎢
⎣
2
..
.
⎥
m
⎥
⎥
(B −1 B̂)·k k ,
⎥ xB +
⎥
k1
⎦
(13.48)
m
where (B −1 B̂)·k denotes the kth column of B −1 B̂ and xB represents the basic solution for
the unperturbed right-hand side b.
From (13.47), we have that for all sufficiently small (but positive), (x B̂ ( ))i > 0 for
all i. Hence, the basis is nondegenerate for the perturbed problem, and we can perform a
step of the simplex method that produces a nonzero (but tiny) decrease in the objective.
Indeed, if we retain the perturbation over all subsequent iterations, and provided
that the initial choice of was small enough, we claim that all subsequent bases visited by
the algorithm are nondegenerate. We prove this claim by contradiction, by assuming that
there is some basis matrix B such that (xB ( ))i 0 for some i and all sufficiently small.
From (13.48), we see that this can happen only when (xB )i 0 and (B −1 B̄)ik 0 for
k 1, 2, . . . , m. The latter relation implies that the ith row of B −1 B̄ is zero, which cannot
occur, because both B and B̄ are nonsingular.
We conclude that, provided the initial choice of is sufficiently small to ensure
nondegeneracy of all subsequent bases, no basis is visited more than once by the simplex
method and therefore, by the same logic as in the proof of Theorem 13.4, the method
terminates finitely at a solution of the perturbed problem. The perturbation can be removed
in a postprocessing phase, by resetting xB B −1 b for the final basis B and the original
right-hand side b.
The question remains of how to choose small enough at the point at which the
original degenerate basis B̂ is encountered. The lexicographic strategy finesses this issue by
not making an explicit choice of , but rather keeping track of the dependence of each basic
variable on each power of . When it comes to selecting the leaving variable, it chooses the
index p that minimizes (xB ( ))i /di over all variables in the basis, for all sufficiently small
. (The choice of p is uniquely defined by this procedure, as we can show by an argument
similar to the one above concerning nondenegeracy of each basis.) We can extend the pivot
procedure slightly to update the dependence of each basic variable on the powers of at
each iteration, including the variable xq that has just entered the basis.
13.6
THE DUAL SIMPLEX METHOD
Here we describe another variant of the simplex method that is useful in a variety of situations
and is often faster on many practical problems than the variant described above. This dual
13.6.
THE DUAL SIMPLEX METHOD
simplex method uses many of the same concepts and methodology described above, such as
the splitting of the matrix A into column submatrices B and N and the generation of iterates
(x, λ, s) that satisfy the complementarity condition x T s 0. The method of Section 13.3
starts with a feasible x (with xB ≥ 0 and xN 0) and a corresponding dual iterate (λ, s)
for which sB 0 but sN is not necessarily nonnegative. After making systematic column
interchanges between B and N , it finally reaches a feasible dual point (λ, s) at which sN ≥ 0,
thus yielding a solution of both the primal problem (13.1) and the dual (13.8). By contrast,
the dual simplex method starts with a point (λ, s) feasible for (13.8), at which sN ≥ 0
and sB 0, and a corresponding primal feasible point x for which xN 0 but xB is not
necessarily nonnegative. By making systematic column interchanges between B and N , it
finally reaches a feasible primal point x for which xB ≥ 0, signifying optimality. Note that
although the matrix B used in this algorithm is a nonsingular column submatrix of A, it
is no longer correct to refer to it as a basis matrix, since it does not satisfy the feasibility
condition xB B −1 b ≥ 0.
We now describe a single step of this method in a similar fashion to Section 13.3,
though the details are a little more complicated here. As mentioned above, we commence
each step with submatrices B and N of A, and corresponding sets B and N . The primal and
dual variables corresponding to these sets are defined as follows (cf. (13.18), (13.20), and
(13.21)):
xB B −1 b,
λB
sB cB − B λ 0,
T
xN 0,
−T
(13.49a)
cB ,
(13.49b)
sN cN − N λ ≥ 0,
T
(13.49c)
If xB ≥ 0, the current point (x, λ, s) satisfies the optimality conditions (13.4), and we are
done. Otherwise, we select a leaving index q ∈ B such that xq < 0. Our aim is to move xq to
zero (thereby ensuring that nonnegativity holds for this component), while allowing sq to
increase away from zero. We will also identify an entering index r ∈ N , such that sr becomes
zero on this step while xr increases away from zero. Hence, the index q will move from B to
N , while r will move from N to B. How do we choose r , and how are x, λ, and s changed
on this step? The description below provides the answer. We use (x + , λ+ , s + ) to denote the
updated values of our variables, after this step is taken.
First, let eq the vector of length m that contains all zeros except for a 1 in the position
occupied by index q in the set B. Since we increase sq away from zero while fixing the
remaining components of sB at zero, the updated value sB+ will have the form
sB+ sB + αeq
(13.50)
for some positive scalar α to be determined. We write the corresponding update to λ as
λ+ λ + αv,
(13.51)
383
384
CHAPTER 13.
THE SIMPLEX METHOD
for some vector v. In fact, since sB+ and λ+ must satisfy the first equation in (13.49c), we
must have
sB+ cB − B T λ+
⇒ sB + αeq cB − B T (λ + αv)
⇒ eq −B T v,
(13.52)
which is a system of equations that we can solve to obtain v.
To see how the dual objective value b T λ changes as a result of this step, we use (13.52)
and the fact that xq xBT eq to obtain
b T λ+ b T λ + αb T v
b T λ − αb T B −T eq
from (13.52)
b λ − αxB eq
from (13.49a)
b λ − αxq
by definition of eq .
T
T
T
Since xq < 0 and since our aim is to maximize the dual objective, we would like to choose α
as large as possible. The upper bound on α is provided by the constraint sN+ ≥ 0. Similarly
to (13.49c), we have
sN+ cN − N T λ+ sN − α N T v sN − αw,
where we have defined
w N T v −N T B −T eq .
The largest α for which sN+ ≥ 0 is given by the formula
α
min
j∈N , w j >0
sj
.
wj
We define the entering index r to be the index at which the minimum in this expression is
achieved. Note that
sr+ 0
and
wr ArT v > 0,
(13.53)
where, as usual, Ar denotes the r th column of A.
Having now identified how λ and s are updated on this step, we need to figure out
how x changes. For the leaving index q, we need to set xq+ 0, while for the entering index
r we can allow xr+ to be nonzero. We denote the direction of change for xB to be the vector
13.7.
PRESOLVING
d, defined by the following linear system:
Bd Ai di Ar .
(13.54)
i∈B
Since from (13.49a), we have
Ai xi b,
i∈B
we have that
Ai (xi − γ di ) + Ar γ b,
(13.55)
i∈B
for any scalar γ . To ensure that xq+ 0, we set
γ xq
,
dq
(13.56)
which is well defined only if dq is nonzero. In fact, we have that dq < 0, since
dq d T eq ArT B −T eq −ArT v −wr < 0,
where we have used the definition of eq along with (13.54), (13.52), and (13.53) to derive
these relationships. Since xq < 0, it follows from (13.56) that γ > 0. Following (13.55) we
can define the updated vector x + as follows:
xi+ 13.7
⎧
xi − γ di ,
⎪
⎪
⎪
⎪
⎨ 0,
⎪
0,
⎪
⎪
⎪
⎩ γ,
for i ∈ B with i q,
for i q,
for i ∈ N with i r ,
for i r .
PRESOLVING
Presolving (also known as preprocessing) is carried out in practical linear programming
codes to reduce the size of the user-defined linear programming problem before passing it
to the solver. A variety of techniques—some obvious, some ingenious—are used to eliminate
certain variables, constraints, and bounds from the problem. Often the reduction in problem
size is quite dramatic, and the linear programming algorithm takes much less time when
applied to the presolved problem than when applied to the original problem. Presolving is
385
386
CHAPTER 13.
THE SIMPLEX METHOD
beneficial regardless of what algorithm is used to solve the linear program; it is used both
in simplex and interior-point codes. Infeasibility may also be detected by the presolver,
eliminating the need to call the linear programming algorithm at all.
We mention just a few of the more straightforward preprocessing techniques here,
referring the interested reader to Andersen and Andersen [4] for a more comprehensive list.
For the purpose of this discussion, we assume that the linear program is formulated with
both lower and upper bounds on x, that is,
min c T x, subject to Ax b, l ≤ x ≤ u,
(13.57)
where some components li of the lower bound vector may be −∞ and some upper bounds
u i may be +∞.
Consider first a row singleton, which happens when one of the equality constraints
involves just one of the variables. Specifically, if constraint k involves only variable j (that
is, Ak j 0, but Aki 0 for all i j), we can immediately set x j bk /Ak j and eliminate
x j from the problem. Note that if this value of x j violates its bounds (that is, x j < l j or
x j > u j ), we can declare the problem to be infeasible, and terminate.
Another obvious technique is the free column singleton, in which there is a variable x j
that occurs in only one of the equality constraints, and is free (that is, its lower bound is −∞
and its upper bound is +∞). In this case, we have for some k that Ak j 0 while Al j 0
for all l k. Here we can simply use constraint k to eliminate x j from the problem, setting
xj bk −
p j
Ak j
Akp x p
.
Once the values of x p for p j have been obtained by solving a reduced linear program,
we can substitute into this formula to recover x j prior to returning the result to the user.
This substitution does not require us to modify any other constraints, but it will change the
cost vector c in general, whenever c j 0. We will need to make the replacement
c p ← c p − c j Akp /Ak j , for all p j.
In this case, we can also determine the dual variable associated with constraint k. Since x j is
a free variable, there is no dual slack associated with it, so the jth dual constraint becomes
m
Al j λl c j ⇒ Ak j λk c j ,
l1
from which we deduce that λk c j /Ak j .
Perhaps the simplest preprocessing check is for the presence of zero rows and columns
in A. If Aki 0 for all i 1, 2, . . . , n, then provided that the right-hand side is also
zero (bk 0), we can simply delete this row from the problem and set the corresponding
13.7.
PRESOLVING
Lagrange multiplier λk to an arbitrary value. For a zero column—say, Ak j 0 for all
k 1, 2, . . . , m—we can determing the optimal value of x j by inspecting its cost coefficient
c j and its bounds l j and u j . If c j < 0, we set x j u j to minimize the product c j x j . (We
are free to do so because x j is not restricted by any of the equality constraints.) If c j < 0
and u j +∞, then the problem is unbounded. Similarly, if c j > 0, we set x j l j , or else
declare unboundedness if l j −∞.
A somewhat more subtle presolving technique is to check for forcing or dominated
constraints. Rather than give a general specification, we illustrate this case with a simple
example. Suppose that one of the equality constraints is as follows:
5x1 − x4 + 2x5 10,
where the variables in question have the following bounds:
0 ≤ x1 ≤ 1, −1 ≤ x4 ≤ 5, 0 ≤ x5 ≤ 2.
It is not hard to see that the equality constraint can only be satisfied if x1 and x5 are at their
upper bounds and x4 is at its lower bound. Any other feasible values of these variables would
result in the left-hand side of the equality constraint being strictly less than 10. Hence, we
can set x1 1, x4 −1, x5 2 and eliminate these variables, and the equality constraint,
from the problem.
We use a similar example to illustrate dominated constraints. Suppose that we have
the following constraint involving three variables:
2x2 + x6 − 3x7 8,
where the variables in question have the following bounds:
−10 ≤ x2 ≤ 10, 0 ≤ x6 ≤ 1, 0 ≤ x7 ≤ 2.
By rearranging the constraint and using the bounds on x6 and x7 , we find that
x2 4 − (1/2)x6 + (3/2)x7 ≤ 4 − 0 + (3/2)2 7.
and similarly, using the opposite bounds on x6 and x7 we obtain x2 ≥ 7/2. We conclude
that the stated bounds of −10 and 10 on x2 are redundant, since x2 is implicitly confined to
an even smaller interval by the combination of the equality constraint and the bounds on
x6 and x7 . Hence, we can drop the bounds on x2 from the formulation and treat it as a free
variable.
Presolving techniques are applied recursively, because the elimination of certain variables or constraints may create situations that allow further eliminations. As a trivial example,
387
388
CHAPTER 13.
THE SIMPLEX METHOD
suppose that the following two equality constraints are present in the problem:
3x2 6, x2 + 4x5 10.
The first of these constraints is a row singleton, which we can use to set x2 2 and
eliminate this variable and constraint. After substitution, the second constraint becomes
4x5 10 − x2 8, which is again a row singleton. We can therefore set x5 2 and
eliminate this variable and constraint as well.
Relatively little information about presolving techniques has appeared in the literature, in part because they have commercial value as an important component of linear
programming software.
13.8
WHERE DOES THE SIMPLEX METHOD FIT?
In linear programming, as in all optimization problems in which inequality constraints are
present, the fundamental task of the algorithm is to determine which of these constraints
are active at the solution (see Definition 12.1 and which are inactive. The simplex method
belongs to a general class of algorithms for constrained optimization known as active set
methods, which explicitly maintain estimates of the active and inactive index sets that are
updated at each step of the algorithm. (At each iteration, the basis B is our current estimate
of the inactive set, that is, the set of indices i for which we suspect that xi > 0 at the
solution of the linear program.) Like most active set methods, the simplex method makes
only modest changes to these index sets at each step; a single index is exchanged between B
into N .
Active set algorithms for quadratic programming, bound-constrained optimization,
and nonlinear programming use the same basic strategy as simplex of making an explicit
estimate of the active set and taking a step toward the solution of a reduced problem in which
the constraints in this estimated active set are satisfied as equalities. When nonlinearity enters
the problem, many of the features that make the simplex method so effective no longer apply.
For example, it is no longer true in general that at least n − m of the bounds x ≥ 0 are active
at the solution, and the specialized linear algebra techniques described in Section 13.5 no
longer apply. Nevertheless, the simplex method is rightly viewed as the antecedent of the
active set class of methods for constrained optimization.
One undesirable feature of the simplex method attracted attention from its earliest
days. Though highly efficient on almost all practical problems (the method generally requires
at most 2m to 3m iterations, where m is the row dimension of the constraint matrix in
(13.1)), there are pathological problems on which the algorithm performs very poorly.
Klee and Minty [182] presented an n-dimensional problem whose feasible polytope has 2n
vertices, for which the simplex method visits every single vertex before reaching the optimal
point! This example verified that the complexity of the simplex method is exponential;
13.8.
WHERE DOES
THE
SIMPLEX METHOD FIT?
roughly speaking, its running time may be an exponential function of the dimension of
the problem. For many years, theoreticians searched for a linear programming algorithm
that has polynomial complexity, that is, an algorithm in which the running time is bounded
by a polynomial function of the amount of storage required to define the problem. In
the late 1970s, Khachiyan [180] described an ellipsoid method that indeed has polynomial
complexity but turned out to be impractical. In the mid-1980s, Karmarkar [175] described
a polynomial algorithm that approaches the solution through the interior of the feasible
polytope rather than working its way around the boundary as the simplex method does.
Karmarkar’s announcement marked the start of intense research in the field of interior-point
methods, which are the subject of the next chapter.
NOTES AND REFERENCES
The standard reference for the simplex method is Dantzig’s book [86]. Later excellent
texts include Chvátal [61] and Vanderbei [293].
Further information on steepest-edge pivoting can be found in Goldfarb and
Reid [134] and Goldfarb and Forrest [133].
An alternative procedure for performing the Phase-I calculation of an initial basis
was described by Wolfe [310]. This technique does not require artificial variables to be
introduced in the problem formulation, but rather starts at any point x that satisfies Ax b
with at most m nonzero components in x. (Note that we do not require the basic part xB to
consist of all positive components.) Phase I then consists in solving the problem
min
x
−xi
subject to Ax b,
xi <0
and terminating when an objective value of 0 is attained. This problem is not a linear
program—its objective is only piecewise linear—but it can be solved by the simplex method
nonetheless. The key is to redefine the cost vector f at each iteration x such that f i −1
for xi < 0 and f i 0 otherwise.
✐
EXERCISES
✐
13.1 Convert the following linear program to standard form:
max c T x + d T y subject to A1 x b1 , A2 x + B2 y ≤ b2 , l ≤ y ≤ u,
x,y
where there are no explicit bounds on x.
✐
13.2 Verify that the dual of (13.8) is the original primal problem (13.1).
389
390
CHAPTER 13.
THE SIMPLEX METHOD
✐
13.3 Complete the proof of Theorem 13.1 by showing that if the dual (13.7) is
unbounded above, the primal (13.1) must be infeasible.
✐
13.4 Theorem 13.1 does not exclude the possibility that both primal and dual are
infeasible. Give a simple linear program for which such is the case.
✐
13.5 Show that the dual of the linear program
min c T x subject to Ax ≥ b, x ≥ 0,
is
max b T λ subject to A T λ ≤ c, λ ≥ 0.
✐
13.6 Show that when m ≤ n and the rows of A are linearly dependent in (13.1), then
the matrix B in (13.13) is singular, and therefore there are no basic feasible points.
✐
13.7 Consider the overdetermined linear system Ax b with m rows and n columns
(m > n). When we apply Gaussian elimination with complete pivoting to A, we obtain
P AQ L
U11
U12
0
0
,
where P and Q are permutation matrices, L is m × m lower triangular, U11 is m̄ × m̄ upper
triangular and nonsingular, U12 is m̄ × (n − m̄), and m̄ ≤ n is the rank of A.
(a) Show that the system Ax b is feasible if the last m − m̄ components of L −1 Pb are
zero, and infeasible otherwise.
(b) When m̄ n, find the unique solution of Ax b.
(c) Show that the reduced system formed from the first m̄ rows of P A and the first m̄
components of Pb is equivalent to Ax b (i.e., a solution of one system also solves
the other).
✐
13.8 Verify formula (13.37).
✐
13.9 Consider the following linear program:
min −5x1 − x2
subject to
x1 + x2 ≤ 5,
2x1 + (1/2)x2 ≤ 8,
x ≥ 0.
13.8.
WHERE DOES
THE
SIMPLEX METHOD FIT?
(a) Add slack variables x3 and x4 to convert this problem to standard form.
(b) Using Procedure 13.1, solve this problem using the simplex method, showing at each
step the basis and the vectors λ, sN , and xB , and the value of the objective function. (The
initial choice of B for which xB ≥ 0 should be obvious once you have added the slacks
in part (a).)
✐
13.10 Calculate the values of l52 , l53 , l54 , and ŵ2 in (13.30), by equating the last row
of L 1 U1 to the last row of the matrix in (13.29).
✐
13.11 By extending the procedure (13.27) appropriately, show how the factorization
(13.31) can be used to solve linear systems with coefficient matrix B + efficiently.
391
This is pag
Printer: O
14
CHAPTER
Linear
Programming:
Interior-Point
Methods
In the 1980s it was discovered that many large linear programs could be solved efficiently by
using formulations and algorithms from nonlinear programming and nonlinear equations.
One characteristic of these methods was that they required all iterates to satisfy the inequality
constraints in the problem strictly, so they became known as interior-point methods. By
the early 1990s, a subclass of interior-point methods known as primal-dual methods had
distinguished themselves as the most efficient practical approaches, and proved to be strong
competitors to the simplex method on large problems. These methods are the focus of this
chapter.
14.1.
PRIMAL-DUAL METHODS
Interior-point methods arose from the search for algorithms with better theoretical
properties than the simplex method. As we mentioned in Chapter 13, the simplex method can
be inefficient on certain pathological problems. Roughly speaking, the time required to solve
a linear program may be exponential in the size of the problem, as measured by the number
of unknowns and the amount of storage needed for the problem data. For almost all practical
problems, the simplex method is much more efficient than this bound would suggest, but
its poor worst-case complexity motivated the development of new algorithms with better
guaranteed performance. The first such method was the ellipsoid method, proposed by
Khachiyan [180], which finds a solution in time that is at worst polynomial in the problem
size. Unfortunately, this method approaches its worst-case bound on all problems and is
not competitive with the simplex method in practice.
Karmarkar’s projective algorithm [175], announced in 1984, also has the polynomial
complexity property, but it came with the added attraction of good practical behavior. The
initial claims of excellent performance on large linear programs were never fully borne
out, but the announcement prompted a great deal of research activity which gave rise to
many new methods. All are related to Karmarkar’s original algorithm, and to the log-barrier
approach described in Chapter 19, but many of the approaches can be motivated and
analyzed independently of the earlier methods.
Interior-point methods share common features that distinguish them from the simplex
method. Each interior-point iteration is expensive to compute and can make significant
progress towards the solution, while the simplex method usually requires a larger number of
inexpensive iterations. Geometrically speaking, the simplex method works its way around
the boundary of the feasible polytope, testing a sequence of vertices in turn until it finds the
optimal one. Interior-point methods approach the boundary of the feasible set only in the
limit. They may approach the solution either from the interior or the exterior of the feasible
region, but they never actually lie on the boundary of this region.
In this chapter, we outline some of the basic ideas behind primal-dual interior-point
methods, including the relationship to Newton’s method and homotopy methods and the
concept of the central path. We sketch the important methods in this class, and give a comprehensive convergence analysis of a particular interior-point method known as a long-step
path-following method. We describe in some detail a practical predictor-corrector algorithm
proposed by Mehrotra, which is the basis of much of the current generation of software.
14.1
PRIMAL-DUAL METHODS
OUTLINE
We consider the linear programming problem in standard form; that is,
min c T x, subject to Ax b, x ≥ 0,
(14.1)
where c and x are vectors in IRn , b is a vector in IRm , and A is an m × n matrix with full row
393
394
CHAPTER 14.
INTERIOR-POINT METHODS
rank. (As in Chapter 13, we can preprocess the problem to remove dependent rows from A
if necessary.) The dual problem for (14.1) is
max b T λ, subject to A T λ + s c, s ≥ 0,
(14.2)
where λ is a vector in IRm and s is a vector in IRn . As shown in Chapter 13, solutions of
(14.1),(14.2) are characterized by the Karush–Kuhn–Tucker conditions (13.4), which we
restate here as follows:
A T λ + s c,
(14.3a)
Ax b,
(14.3b)
xi si 0, i 1, 2, . . . , n,
(14.3c)
(x, s) ≥ 0.
(14.3d)
Primal-dual methods find solutions (x ∗ , λ∗ , s ∗ ) of this system by applying variants of
Newton’s method to the three equalities in (14.3) and modifying the search directions and
step lengths so that the inequalities (x, s) ≥ 0 are satisfied strictly at every iteration. The
equations (14.3a), (14.3b), (14.3c) are linear or only mildly nonlinear and so are not difficult
to solve by themselves. However, the problem becomes much more difficult when we add the
nonnegativity requirement (14.3d), which gives rise to all the complications in the design
and analysis of interior-point methods.
To derive primal-dual interior-point methods we restate the optimality conditions
(14.3) in a slightly different form by means of a mapping F from IR2n+m to IR2n+m :
⎡
⎢
F(x, λ, s) ⎣
AT λ + s − c
Ax − b
⎤
⎥
⎦ 0,
(14.4a)
(x, s) ≥ 0,
(14.4b)
X Se
where
X diag(x1 , x2 , . . . , xn ),
S diag(s1 , s2 , . . . , sn ),
(14.5)
and e (1, 1, . . . , 1)T . Primal-dual methods generate iterates (x k , λk , s k ) that satisfy the
bounds (14.4b) strictly, that is, x k > 0 and s k > 0. This property is the origin of the term
interior-point. By respecting these bounds, the methods avoid spurious solutions, that is,
points that satisfy F(x, λ, s) 0 but not (x, s) ≥ 0. Spurious solutions abound, and do not
provide useful information about solutions of (14.1) or (14.2), so it makes sense to exclude
them altogether from the region of search.
Like most iterative algorithms in optimization, primal-dual interior-point methods
have two basic ingredients: a procedure for determining the step and a measure of the
14.1.
PRIMAL-DUAL METHODS
desirability of each point in the search space. An important component of the measure of
desirability is the average value of the pairwise products xi si , i 1, 2, . . . , n, which are all
positive when x > 0 and s > 0. This quantity is known as the duality measure and is defined
as follows:
µ
n
xT s
1
.
xi si n i1
n
(14.6)
The procedure for determining the search direction has its origins in Newton’s method
for the nonlinear equations (14.4a). Newton’s method forms a linear model for F around
the current point and obtains the search direction (x, λ, s) by solving the following
system of linear equations:
⎡
x
⎤
⎢
⎥
J (x, λ, s) ⎣ λ ⎦ −F(x, λ, s),
s
where J is the Jacobian of F. (See Chapter 11 for a detailed discussion of Newton’s method
for nonlinear systems.) If we use the notation rc and rb for the first two block rows in F,
that is,
rb Ax − b,
rc A T λ + s − c,
(14.7)
we can write the Newton equations as follows:
⎡
0
⎢
⎣ A
S
AT
I
⎤⎡
x
⎤
⎡
−rc
0
⎥⎢
⎥ ⎢
0 ⎦ ⎣ λ ⎦ ⎣
−rb
0
X
s
−X Se
⎤
⎥
⎦.
(14.8)
Usually, a full step along this direction would violate the bound (x, s) ≥ 0, so we perform a
line search along the Newton direction and define the new iterate as
(x, λ, s) + α(x, λ, s),
for some line search parameter α ∈ (0, 1]. We often can take only a small step along this
direction (α 1) before violating the condition (x, s) > 0. Hence, the pure Newton
direction (14.8), sometimes known as the affine scaling direction, often does not allow us to
make much progress toward a solution.
Most primal-dual methods use a less aggressive Newton direction, one that does not
aim directly for a solution of (14.3a), (14.3b), (14.3c) but rather for a point whose pairwise
products xi si are reduced to a lower average value—not all the way to zero. Specifically, we
395
396
CHAPTER 14.
INTERIOR-POINT METHODS
take a Newton step toward the a point for which xi si σ µ, where µ is the current duality
measure and σ ∈ [0, 1] is the reduction factor that we wish to achieve in the duality measure
on this step. The modified step equation is then
⎡
0
⎢
⎣ A
S
AT
I
⎤⎡
x
⎤
⎡
−rc
0
⎥ ⎢
⎥⎢
0 ⎦ ⎣ λ ⎦ ⎣
−rb
0
X
s
−X Se + σ µe
⎤
⎥
⎦.
(14.9)
We call σ the centering parameter, for reasons to be discussed below. When σ > 0, it usually
is possible to take a longer step α along the direction defined by (14.16) before violating the
bounds (x, s) ≥ 0.
At this point, we have specified most of the elements of a path-following primal-dual
interior-point method. The general framework for such methods is as follows.
Framework 14.1 (Primal-Dual Path-Following).
Given (x 0 , λ0 , s 0 ) with (x 0 , s 0 ) > 0;
for k 0, 1, 2, . . .
Choose σk ∈ [0, 1] and solve
⎡
0
⎢
⎣ A
S
k
AT
⎤⎡
I
x k
⎤
⎡
−rck
⎤
⎥
⎦,
(14.10)
(x k+1 , λk+1 , s k+1 ) (x k , λk , s k ) + αk (x k , λk , s k ),
(14.11)
0
0
⎥⎢
⎥ ⎢
0 ⎦ ⎣ λk ⎦ ⎣
−rbk
s
−X S e + σk µk e
X
k
k
k k
where µk (x k )T s k /n;
Set
choosing αk so that (x k+1 , s k+1 ) > 0.
end (for).
The choices of centering parameter σk and step length αk are crucial to the performance
of the method. Techniques for controlling these parameters, directly and indirectly, give rise
to a wide variety of methods with diverse properties.
Although software for implementing interior-point methods does not usually start
from a point (x 0 , λ0 , s 0 ) that is feasible with respect to the linear equations (14.3a) and
(14.3b), most of the historical development of theory and algorithms assumed that these
conditions are satisfied. In the remainder of this section, we discuss this feasible case,
showing that a comprehensive convergence analysis can be presented in just a few pages,
using only basic mathematical tools and concepts. Analysis of the infeasible case follows the
14.1.
PRIMAL-DUAL METHODS
same principles, but is considerably more complicated in the details, so we do not present
it here. In Section 14.2, however, we describe a complete practical algorithm that does not
require starting from a feasible initial point.
To begin our discussion and analysis of feasible interior-point methods, we introduce
the concept of the central path, and then describe neighborhoods of this path.
THE CENTRAL PATH
The primal-dual feasible set F and strictly feasible set F o are defined as follows:
F {(x, λ, s) | Ax b, A T λ + s c, (x, s) ≥ 0},
(14.12a)
F {(x, λ, s) | Ax b, A λ + s c, (x, s) > 0}.
(14.12b)
o
T
The central path C is an arc of strictly feasible points that plays a vital role in primal-dual
algorithms. It is parametrized by a scalar τ > 0, and each point (xτ , λτ , sτ ) ∈ C satisfies the
following equations:
A T λ + s c,
(14.13a)
Ax b,
(14.13b)
xi si τ,
i 1, 2, . . . , n,
(x, s) > 0.
(14.13c)
(14.13d)
These conditions differ from the KKT conditions only in the term τ on the right-hand side
of (14.13c). Instead of the complementarity condition (14.3c), we require that the pairwise
products xi si have the same (positive) value for all indices i. From (14.13), we can define
the central path as
C {(xτ , λτ , sτ ) | τ > 0}.
It can be shown that (xτ , λτ , sτ ) is defined uniquely for each τ > 0 if and only if F o is
nonempty.
The conditions (14.13) are also the optimality conditions for a logarithmic-barrier
formulation of the problem (14.1). By introducing log-barrier terms for the nonnegativity
constraints, with barrier parameter τ > 0, we obtain
min c T x − τ
n
i1
ln xi , subject to Ax b.
(14.14)
397
398
CHAPTER 14.
INTERIOR-POINT METHODS
The KKT conditions (12.34) for this problem, with Lagrange multiplier λ for the equality
constraint, are as follows:
ci −
τ
− A·iT λ, i 1, 2, . . . , n, Ax b.
xi
Since the objective in (14.14) is strictly convex, these conditions are sufficient as well as
necessary for optimality. We recover (14.13) by defining si τ/xi , i 1, 2, . . . , n.
Another way of defining C is to use the mapping F defined in (14.4) and write
⎡
0
⎤
⎢
⎥
F(xτ , λτ , sτ ) ⎣ 0 ⎦ ,
(xτ , sτ ) > 0.
(14.15)
τe
The equations (14.13) approximate (14.3) more and more closely as τ goes to zero. If
C converges to anything as τ ↓ 0, it must converge to a primal-dual solution of the linear
program. The central path thus guides us to a solution along a route that maintains positivity
of the x and s components and decreases the pairwise products xi si , i 1, 2, . . . , n to zero
at the same rate.
Most primal-dual algorithms take Newton steps toward points on C for which τ > 0,
rather than pure Newton steps for F. Since these steps are biased toward the interior of
the nonnegative orthant defined by (x, s) ≥ 0, it usually is possible to take longer steps
along them than along the pure Newton (affine scaling) steps, before violating the positivity
condition.
In the feasible case of (x, λ, s) ∈ F, we have rb 0 and rc 0, so the search direction
satisfies a special case of (14.8), that is,
⎡
0
⎢
⎣ A
S
AT
I
⎤⎡
x
⎤
⎡
0
⎥⎢
⎥ ⎢
0 ⎦ ⎣ λ ⎦ ⎣
0
X
s
0
0
⎤
⎥
⎦,
(14.16)
−X Se + σ µe
where µ is the duality measure defined by (14.6) and σ ∈ [0, 1] is the centering parameter. When σ 1, the equations (14.16) define a centering direction, a Newton step
toward the point (xµ , λµ , sµ ) ∈ C, at which all the pairwise products xi si are identical
to the current average value of µ. Centering directions are usually biased strongly toward
the interior of the nonnegative orthant and make little, if any, progress in reducing the
duality measure µ. However, by moving closer to C, they set the scene for a substantial
reduction in µ on the next iteration. At the other extreme, the value σ 0 gives the
standard Newton (affine scaling) step. Many algorithms use intermediate values of σ from
14.1.
PRIMAL-DUAL METHODS
the open interval (0, 1) to trade off between the twin goals of reducing µ and improving
centrality.
CENTRAL PATH NEIGHBORHOODS AND PATH-FOLLOWING METHODS
Path-following algorithms explicitly restrict the iterates to a neighborhood of the
central path C and follow C to a solution of the linear program. By preventing the iterates
from coming too close to the boundary of the nonnegative orthant, they ensure that it is
possible to take a nontrivial step along each search direction. Mopreover, by forcing the
duality measure µk to zero as k → ∞, we ensure that the iterates (x k , λk , s k ) come closer
and closer to satisfying the KKT conditions (14.3).
The two most interesting neighborhoods of C are
N2 (θ ) {(x, λ, s) ∈ F o | X Se − µe2 ≤ θ µ},
(14.17)
for some θ ∈ [0, 1), and
N−∞ (γ ) {(x, λ, s) ∈ F o | xi si ≥ γ µ all i 1, 2, . . . , n},
(14.18)
for some γ ∈ (0, 1]. (Typical values of the parameters are θ 0.5 and γ 10−3 .) If
a point lies in N−∞ (γ ), each pairwise product xi si must be at least some small multiple
γ of their average value µ. This requirement is actually quite modest, and we can make
N−∞ (γ ) encompass most of the feasible region F by choosing γ close to zero. The N2 (θ )
neighborhood is more restrictive, since certain points in F o do not belong to N2 (θ ) no
matter how close θ is chosen to its upper bound of 1.
By keeping all iterates inside one or other of these neighborhoods, path-following
methods reduce all the pairwise products xi si to zero at more or less the same rate. Figure 14.1
shows the projection of the central path C onto the primal variables for a typical problem,
along with a typical neighborhood N .
Path-following methods are akin to homotopy methods for general nonlinear equations, which also define a path to be followed to the solution. Traditional homotopy methods
stay in a tight tubular neighborhood of their path, making incremental changes to the parameter and chasing the homotopy path all the way to a solution. For primal-dual methods,
this neighborhood is horn-shaped rather than tubular, and it tends to be broad and loose
for larger values of the duality measure µ. It narrows as µ → 0, however, because of the
positivity requirement (x, s) > 0.
The algorithm we specify below, a special case of Framework 14.1, is known as a
long-step path-following algorithm. This algorithm can make rapid progress because of its
use of the wide neighborhood N−∞ (γ ), for γ close to zero. It depends on two parameters
σmin and σmax , which are lower and upper bounds on the centering parameter σk . The search
direction is, as usual, obtained by solving (14.10), and we choose the step length αk to be as
large as possible, subject to the requirement that we stay inside N−∞ (γ ).
399
400
CHAPTER 14.
INTERIOR-POINT METHODS
central path neighborhood
C
Figure 14.1 Central path, projected into space of primal variables x, showing a
typical neighborhood N .
Here and in later analysis, we use the notation
def
(x k (α), λk (α), s k (α)) (x k , λk , s k ) + α(x k , λk , s k ),
def
µk (α) x (α) s (α)/n.
k
T k
(14.19a)
(14.19b)
Algorithm 14.2 (Long-Step Path-Following).
Given γ , σmin , σmax with γ ∈ (0, 1), 0 < σmin ≤ σmax < 1,
and (x 0 , λ0 , s 0 ) ∈ N−∞ (γ );
for k 0, 1, 2, . . .
Choose σk ∈ [σmin , σmax ];
Solve (14.10) to obtain (x k , λk , s k );
Choose αk as the largest value of α in [0, 1] such that
(x k (α), λk (α), s k (α)) ∈ N−∞ (γ );
(14.20)
Set (x k+1 , λk+1 , s k+1 ) (x k (αk ), λk (αk ), s k (αk ));
end (for).
Typical behavior of the algorithm is illustrated in Figure 14.2 for the case of n 2.
The horizontal and vertical axes in this figure represent the pairwise products x1 s1 and x2 s2 ,
so the central path C is the line emanating from the origin at an angle of 45◦ . (A point at the
origin of this illustration is a primal-dual solution if it also satisfies the feasibility conditions
14.1.
x2 s2
1
iterates
PRIMAL-DUAL METHODS
0
central path C
2
3
boundary of neighborhood N
x1 s1
Figure 14.2 Iterates of Algorithm 14.2, plotted in (xs) space.
(14.3a), (14.3b), and (14.3d).) In the unusual geometry of Figure 14.2, the search directions
(x k , λk , s k ) transform to curves rather than straight lines.
As Figure 14.2 shows (and the analysis confirms), the lower bound σmin on the
centering parameter ensures that each search direction starts out by moving away from the
boundary of N−∞ (γ ) and into the relative interior of this neighborhood. That is, small
steps along the search direction improve the centrality. Larger values of α take us outside
the neighborhood again, since the error in approximating the nonlinear system (14.15) by
the linear step equations (14.16) becomes more pronounced as α increases. Still, we are
guaranteed that a certain minimum step can be taken before we reach the boundary of
N−∞ (γ ), as we show in the analysis below.
The analysis of Algorithm 14.2 appears in the next few pages. With judicious choices
of σk , this algorithm is fairly efficient in practice. With a few more modifications, it becomes
the basis of a truly competitive method, as we discuss in Section 14.2.
Our aim in the analysis below is to show that given some small tolerance > 0, the
algorithm requires O(n| log |) iterations to reduce the duality measure by a factor of , that
is, to identify a point (x k , λk , s k ) for which µk ≤ µ0 . For small , the point (x k , λk , s k )
satisfies the primal-dual optimality conditions except for perturbations of about in the
right-hand side of (14.3c), so it is usually very close to a primal-dual solution of the
original linear program. The O(n| log |) estimate is a worst-case bound on the number
of iterations required; on practical problems, the number of iterations required appears
401
402
CHAPTER 14.
INTERIOR-POINT METHODS
to increase only slightly (if at all) as n increases. The simplex method may require 2n
iterations to solve a problem with n variables, though in practice it usually requires a
modest multiple of m iterations, where m is the row dimension of the constraint matrix
A in (14.1).
As is typical for interior-point methods, the analysis builds from a purely technical lemma to a powerful theorem in just a few pages. We start with the technical result
(Lemma 14.1) and use it to derive a bound on the vector of pairwise products xi si ,
i 1, 2, . . . , n (Lemma 14.2). Theorem 14.3 finds a lower bound on the step length αk and
a corresponding estimate of the reduction in µ on iteration k. Finally, Theorem 14.4 proves
that O(n| log |) iterations are required to identify a point for which µk < , for a given
tolerance ∈ (0, 1).
Lemma 14.1.
Let u and v be any two vectors in IRn with u T v ≥ 0. Then
U V e2 ≤ 2−3/2 u + v22 ,
where
U diag(u 1 , u 2 , . . . , u n ),
V diag(v1 , v2 , . . . , vn ).
PROOF. (When the subscript is omitted from · , we mean · 2 , as is our convention
throughout the book.) First, note that for any two scalars α and β with αβ ≥ 0, we have
from the algebraic-geometric mean inequality that
1
|αβ| ≤ |α + β|.
2
(14.21)
Since u T v ≥ 0, we have
0 ≤ uT v u i vi ≥0
u i vi +
u i vi u i vi <0
i∈P
|u i vi | −
i∈M
where we partitioned the index set {1, 2, . . . , n} as
P {i | u i vi ≥ 0},
M {i | u i vi < 0}.
|u i vi |,
(14.22)
14.1.
PRIMAL-DUAL METHODS
Now,
1/2
U V e [u i vi ]i∈P 2 + [u i vi ]i∈M 2
1/2
≤ [u i vi ]i∈P 21 + [u i vi ]i∈M 21
1/2
≤ 2 [u i vi ]i∈P 21
√ 1
2
≤ 2 (u i + vi )
4
i∈P 1
2−3/2
(u i + vi )2
since · 2 ≤ · 1
from (14.22)
from (14.21)
i∈P
n
≤ 2−3/2
(u i + vi )2
i1
≤ 2−3/2 u + v2 ,
completing the proof.
For the next result, we omit the iteration counter k from (14.10), and define the
diagonal matrices X and S similarly to (14.5), as follows:
X diag(x1 , x2 , . . . , xn ),
S diag(s1 , s2 , . . . , sn )..
Lemma 14.2.
If (x, λ, s) ∈ N−∞ (γ ), then
X Se ≤ 2−3/2 (1 + 1/γ )nµ.
PROOF. It is easy to show using (14.10) that
x T s 0.
(14.23)
By multiplying the last block row in (14.10) by (X S)−1/2 and using the definition D X 1/2 S −1/2 , we obtain
D −1 x + Ds (X S)−1/2 (−X Se + σ µe).
(14.24)
Because (D −1 x)T (Ds) x T s 0, we can apply Lemma 14.1 with u D −1 x
and v Ds to obtain
X Se (D −1 X )(DS)e
≤ 2−3/2 D −1 x + Ds2
−3/2
2
(X S)
−1/2
from Lemma 14.1
(−X Se + σ µe)
2
from (14.24).
Expanding the squared Euclidean norm and using such relationships as x T s nµ and
403
404
CHAPTER 14.
INTERIOR-POINT METHODS
e T e n, we obtain
n
1
X Se ≤ 2
x s − 2σ µe e + σ µ
xs
i1 i i
n
≤ 2−3/2 x T s − 2σ µe T e + σ 2 µ2
γµ
2
σ
nµ
≤ 2−3/2 1 − 2σ +
γ
−3/2
T
T
2
2
since xi si ≥ γ µ
≤ 2−3/2 (1 + 1/γ )nµ,
as claimed.
Theorem 14.3.
Given the parameters γ , σmin , and σmax in Algorithm 14.2, there is a constant δ
independent of n such that
δ
µk ,
µk+1 ≤ 1 −
n
(14.25)
for all k ≥ 0.
PROOF. We start by proving that
1 − γ σk
x k (α), λk (α), s k (α) ∈ N−∞ (γ ) for all α ∈ 0, 23/2 γ
,
1+γ n
(14.26)
where x k (α), λk (α), s k (α) is defined as in (14.19). It follows that the step length αk is at
least as long as the upper bound of this interval, that is,
αk ≥ 23/2
σk 1 − γ
γ
.
n 1+γ
(14.27)
For any i 1, 2, . . . , n, we have from Lemma 14.2 that
|xik sik | ≤ X k S k e2 ≤ 2−3/2 (1 + 1/γ )nµk .
Using (14.10), we have from xik sik ≥ γ µk and (14.28) that
xik (α)sik (α) xik + αxik sik + αsik
xik sik + α xik sik + sik xik + α 2 xik sik
≥ xik sik (1 − α) + ασk µk − α 2 |xik sik |
≥ γ (1 − α)µk + ασk µk − α 2 2−3/2 (1 + 1/γ )nµk .
(14.28)
14.1.
PRIMAL-DUAL METHODS
By summing the n components of the equation S k x k + X k s k −X k S k e + σk µk e (the
third block row from (14.10)), and using (14.23) and the definition of µk and µk (α) (see
(14.19)), we obtain
µk (α) (1 − α(1 − σk ))µk .
From these last two formulas, we can see that the proximity condition
xik (α)sik (α) ≥ γ µk (α)
is satisfied, provided that
γ (1 − α)µk + ασk µk − α 2 2−3/2 (1 + 1/γ )nµk ≥ γ (1 − α + ασk )µk .
Rearranging this expression, we obtain
ασk µk (1 − γ ) ≥ α 2 2−3/2 nµk (1 + 1/γ ),
which is true if
α≤
23/2
1−γ
σk γ
.
n
1+γ
We have proved that x k (α), λk (α), s k (α) satisfies the proximity condition for N−∞ (γ )
when α lies in the range stated in (14.26). It is not difficult to show that
k
x (α), λk (α), s k (α) ∈ F o for all α in the given range. Hence, we have proved (14.26)
and therefore (14.27).
We complete the proof of the theorem by estimating the reduction in µ on the kth
step. Because of (14.23), (14.27), and the last block row of (14.16), we have
µk+1 x k (αk )T s k (αk )/n
(x k )T s k + αk (x k )T s k + (s k )T x k + αk2 (x k )T s k /n
µk + αk −(x k )T s k /n + σk µk
(1 − αk (1 − σk ))µk
23/2 1 − γ
γ
σk (1 − σk ) µk .
≤ 1−
n 1+γ
(14.29)
Now, the function σ (1 − σ ) is a concave quadratic function of σ , so on any given interval
it attains its minimum value at one of the endpoints. Hence, we have
σk (1 − σk ) ≥ min {σmin (1 − σmin ), σmax (1 − σmax )} , for all σk ∈ [σmin , σmax ].
405
406
CHAPTER 14.
INTERIOR-POINT METHODS
The proof is completed by substituting this estimate into (14.29) and setting
δ 23/2 γ
1−γ
min {σmin (1 − σmin ), σmax (1 − σmax )} .
1+γ
We conclude with a result that a reduction of a factor of
can be obtained in O(n log 1/ ) iterations.
in the duality measure µ
Theorem 14.4.
Given ∈ (0, 1) and γ ∈ (0, 1), suppose the starting point in Algorithm 14.2 satisfies
(x 0 , λ0 , s 0 ) ∈ N−∞ (γ ). Then there is an index K with K O(n log 1/ ) such that
µk ≤ µ0 ,
for all k ≥ K .
PROOF. By taking logarithms of both sides in (14.25), we obtain
log µk+1
δ
+ log µk .
≤ log 1 −
n
By applying this formula repeatedly, we have
δ
+ log µ0 .
log µk ≤ k log 1 −
n
The following well-known estimate for the log function,
log(1 + β) ≤ β,
for all β > −1,
implies that
δ
.
log(µk /µ0 ) ≤ k −
n
Therefore, the condition µk /µ0 ≤ is satisfied if we have
δ
≤ log .
k −
n
This inequality holds for all k that satisfy
def
k≥K so the proof is complete.
1
n
n
log | log |,
δ
δ
14.2.
14.2
PRACTICAL PRIMAL-DUAL ALGORITHMS
PRACTICAL PRIMAL-DUAL ALGORITHMS
Practical implementations of interior-point algorithms follow the spirit of the previous
section, in that strict positivity of x k and s k is maintained throughout and each step is a
Newton-like step involving a centering component. However, most implementations work
with an infeasible starting point and infeasible iterations. Several aspects of “theoretical”
algorithms are typically ignored, while several enhancements are added that have a significant
effect on practical performance. In this section, we describe the algorithmic enhancements
that are found in a typical implementation of an infeasible-interior-point method, and
present the resulting method as Algorithm 14.3. Many of the techniques of this section are
described in the paper of Mehrotra [207], which can be consulted for further details.
CORRECTOR AND CENTERING STEPS
A key feature of practical algorithms is their use of corrector steps that compensate for
the linearization error made by the Newton (affine-scaling) step in modeling the equation
xi si 0, i 1, 2, . . . , n (see (14.3c)). Consider the affine-scaling direction (x, λ, s)
defined by
⎡
0
⎢
⎣ A
AT
0
0
S
I
⎤⎡
x aff
⎤
⎡
⎥ ⎢
⎥⎢
0 ⎦ ⎣ λaff ⎦ ⎣
s
X
−rc
−rb
⎤
⎥
⎦,
(14.30)
−X Se
aff
(where rb and rc are defined in (14.7)). If we take a full step in this direction, we obtain
(xi + xiaff )(si + siaff )
xi si + xi siaff + si xiaff + xiaff siaff xiaff siaff .
That is, the updated value of xi si is xiaff siaff rather than the ideal value 0. We can solve
the following system to obtain a step (x cor , λcor , s cor ) that attempts to correct for this
deviation from the ideal:
⎡
0
⎢
⎣ A
S
AT
0
0
I
⎤⎡
x cor
⎤
⎡
⎥ ⎢
⎥⎢
0 ⎦ ⎣ λcor ⎦ ⎣
X
s
cor
⎤
0
⎥
⎦.
0
(14.31)
−X S e
aff
aff
In many cases, the combined step (x aff , λaff , s aff ) + (x cor , λcor , s cor ) does a better
job of reducing the duality measure than does the affine-scaling step alone.
Like theoretical algorithms such as the one analysed in Section 14.1, practical algorithms make use of centering steps, with an adaptive choice of the centering parameter σk .
The affine-scaling step can be used as the basis of a successful heuristic for choosing σk .
407
408
CHAPTER 14.
INTERIOR-POINT METHODS
Roughly speaking, if the affine-scaling step (multiplied by a steplength to maintain nonnegativity of x and s) reduces the duality measure significantly, there is not much need for
centering, so a smaller value of σk is appropriate. Conversely, if not much progress can be
made along this direction before reaching the boundary of the nonnegative orthant, a larger
value of σk will ensure that the next iterate is more centered, so a longer step will be possible
from this next point. Specifically, this scheme calculates the maximum allowable steplengths
along the affine-scaling direction (14.30) as follows:
xi
min 1, min − aff ,
xi
i:xiaff <0
si
def
min 1, min − aff ,
si
i:siaff <0
pri def
αaff dual
αaff
(14.32a)
(14.32b)
and then defines µaff to be the value of µ that would be obtained by using these steplengths,
that is,
pri
dual
s aff )/n.
µaff (x + αaff x aff )T (s + αaff
(14.33)
The centering parameter σ is chosen according to the following heuristic (which does not
have a solid analytical justification, but appears to work well in practice):
σ µaff
µ
3
.
(14.34)
To summarize, computation of the search direction requires the solution of two linear
systems. First, the system (14.30) is solved to obtain the affine-scaling direction, also known
as the predictor step. This step is used to define the right-hand side for the corrector step (see
(14.31)) and to calculate the centering parameter from (14.33), (14.34). Second, the search
direction is calculated by solving
⎡
0
⎢
⎣ A
S
AT
I
⎤⎡
x
⎤
⎡
⎤
−rc
0
⎥⎢
⎥ ⎢
0 ⎦ ⎣ λ ⎦ ⎣
−rb
0
s
−X Se − X S e + σ µe
X
aff
⎥
⎦.
(14.35)
aff
Note that the predictor, corrector, and centering contributions have been aggregated on the
right-hand side of this system. The coefficient matrix in both linear systems (14.30) and
(14.35) is the same. Thus, the factorization of the matrix needs to be computed only once,
and the marginal cost of solving the second system is relatively small.
14.2.
PRACTICAL PRIMAL-DUAL ALGORITHMS
STEP LENGTHS
Practical implementations typically do not enforce membership of the central path
neighborhoods N2 and N−∞ defined in the previous section. Rather, they calculate the
maximum steplengths that can be taken in the x and s variables (separately) without
violating nonnegativity, then take a steplength of slightly less than this maximum (but no
greater than 1). Given an iterate (x k , λk , s k ) with (x k , s k ) > 0, and a step (x k , λk , s k ),
pri
dual
defined as follows:
it is easy to show that the quantities αk,max and αk,max
pri
def
αk,max min −
i:xik <0
xik
,
xik
def
dual
αk,max
min −
i:sik <0
sik
,
sik
(14.36)
are the largest values of α for which x k + αx k ≥ 0 and s k + αs k ≥ 0, respectively. (Note
that these formulae are similar to the ratio test used in the simplex method to determine
the index that enters the basis.) Practical algorithms then choose the steplengths to lie in the
open intervals defined by these maxima, that is,
pri
pri
αk ∈ (0, αk,max ),
dual
αkdual ∈ (0, αk,max
),
and then obtain a new iterate by setting
pri
x k+1 x k + αk x k ,
(λk+1 , s k+1 ) (λk , s k ) + αkdual (λk , s k ).
If the step (x k , λk , s k ) rectifies the infeasibility in the KKT conditions (14.3a) and
(14.3b), that is,
Ax k −rbk −(Ax k − b),
A T λk + s k −rck −(A T λk + s k − c),
it is easy to show that the infeasibilities at the new iterate satisfy
pri rbk+1 1 − αk rbk ,
rck+1 1 − αkdual rck .
(14.37)
The following formula is used to calculate steplengths in many practical
implementations
pri
pri
αk min(1, ηk αk,max ),
dual
αkdual min(1, ηk αk,max
),
(14.38)
where ηk ∈ [0.9, 1.0) is chosen so that ηk → 1 as the iterates approach the primal-dual
solution, to accelerate the asymptotic convergence.
409
410
CHAPTER 14.
INTERIOR-POINT METHODS
STARTING POINT
Choice of starting point is an important practical issue with a significant effect on
the robustness of the algorithm. A poor choice (x 0 , λ0 , s 0 ) satisfying only the minimal
conditions x 0 > 0 and s 0 > 0 often leads to failure of convergence. We describe here a
heuristic that finds a starting point that satisfies the equality constraints in the primal and
dual problems reasonably well, while maintaining positivity of the x and s components and
avoiding excessively large values of these components.
First, we find a vector x̃ of minimum norm satisfying the primal constraint Ax b,
and a vector (λ̃, s̃) satisfy the dual constraint A T λ + s c such that s̃ has minimum norm.
That is, we solve the problems
min 12 x T x subject to Ax b,
(14.39a)
min 12 s T s subject to A T λ + s c.
(14.39b)
x
(λ,s)
It is not difficult to show that x̃ and (λ̃, s̃) can be written explicitly as follows:
x̃ A T (A A T )−1 b,
λ̃ (A A T )−1 Ac,
s̃ c − A T λ̃.
(14.40)
In general, x̃ and s̃ will have nonpositive components, so are not suitable for use as a
starting point. We define
δx max(−(3/2) min x̃i , 0),
i
δs max(−(3/2) min s̃i , 0),
i
and adjust the x and s vectors as follows:
x̂ x̃ + δx e,
ŝ s̃ + δs e,
where, as usual, e (1, 1, . . . , 1)T . Clearly, we have x̂ ≥ 0 and ŝ ≥ 0. To ensure that the
components of x 0 and s 0 are not too close to zero and not too dissimilar, we add two more
scalars defined as follows:
δ̂x 1 x̂ T ŝ
,
2 e T ŝ
δ̂s 1 x̂ T ŝ
2 e T x̂
(Note that δ̂x is the average size of the components of x̂, weighted by the corresponding
components of ŝ; similarly for δ̂s .) Finally, we define the starting point as follows:
x 0 x̂ + δ̂x e, λ0 λ̃, s 0 ŝ + δ̂s e.
The computational cost of finding (x 0 , λ0 , s 0 ) by this scheme is about the same as one step
of the primal-dual method.
14.2.
PRACTICAL PRIMAL-DUAL ALGORITHMS
In some cases, we have prior knowledge about the solution, possibly in the form
of a solution to a similar linear program. The use of such “warm-start” information in
constructing a starting point is discussed in Section 14.4.
A PRACTICAL ALGORITHM
We now give a formal specification of a practical algorithm.
Algorithm 14.3 (Predictor-Corrector Algorithm (Mehrotra [207])).
Calculate (x 0 , λ0 , s 0 ) as described above;
for k 0, 1, 2, . . .
Set (x, λ, s) (x k , λk , s k ) and solve (14.30) for (x aff , λaff , s aff );
pri
dual
, and µaff as in (14.32) and (14.33);
Calculate αaff , αaff
Set centering parameter to σ (µaff /µ)3 ;
Solve (14.35) for (x, λ, s);
pri
Calculate αk and αkdual from (14.38);
Set
pri
x k+1 x k + αk x,
(λk+1 , s k+1 ) (λk , s k ) + αkdual (λ, s);
end (for).
No convergence theory is available for Mehrotra’s algorithm, at least in the form in
which it is described above. In fact, there are examples for which the algorithm diverges.
Simple safeguards could be incorporated into the method to force it into the convergence
framework of existing methods or to improve its robustness, but many practical codes do
not implement these safeguards, because failures are rare.
When presented with a linear program that is infeasible or unbounded, the algorithm
above typically diverges, with the infeasibilities rbk and rck and/or the duality measure µk
going to ∞. Since the symptoms of infeasibility and unboundedness are fairly easy to
recognize, interior-point codes contain heuristics to detect and report these conditions.
More rigorous approaches for detecting infeasibility and unboundedness make use of the
homogeneous self-dual formulation; see Wright [316, Chapter 9] and the references therein
for a discussion. A more recent approach that applies directly to infeasible-interior-point
methods is described by Todd [286].
SOLVING THE LINEAR SYSTEMS
Most of the computational effort in primal-dual methods is taken up in solving linear
systems such as (14.9), (14.30), and (14.35). The coefficient matrix in these systems is usually
large and sparse, since the constraint matrix A is itself large and sparse in most applications.
411
412
CHAPTER 14.
INTERIOR-POINT METHODS
The special structure in the step equations allows us to reformulate them as systems with
more compact symmetric coefficient matrices, which are easier and cheaper to factor than
the original sparse form.
We apply the reformulation procedures to the following general form of the linear
system:
⎡
0
⎢
⎣ A
S
AT
I
⎤⎡
x
⎤
⎡
−rc
⎤
0
⎥⎢
⎥ ⎢
⎥
0 ⎦ ⎣ λ ⎦ ⎣ −rb ⎦ .
0
X
s
(14.41)
−r xs
Since x and s are strictly positive, the diagonal matrices X and S are nonsingular. We can
eliminate s and add −X −1 times the third equation in this system to the first equation to
obtain
−D −2 A T
x
−rc + X −1r xs
,
(14.42a)
A
0
λ
−rb
s −X −1 r xs − X −1 Sx,
(14.42b)
where we have introduced the notation
D S −1/2 X 1/2 .
(14.43)
This form of the step equations usually is known as the augmented system. We can go further
and eliminate x and add AD 2 times the first equation to the second equation in (14.42a)
to obtain
AD 2 A T λ −rb − AX S −1 rc + AS −1r xs
(14.44a)
s −rc − A T λ,
(14.44b)
x −S −1 r xs − X S −1 s,
(14.44c)
where the expressions for s and s are obtained from the original system (14.41). The
form (14.44a) often is called the normal-equations form, because the system (14.44a) can
be viewed as the normal equations (10.14) for a certain linear least-squares problem with
coefficient matrix D A T .
Most implementations of primal-dual methods are based on formulations like (14.44).
They use direct sparse Cholesky algorithms to factor the matrix AD 2 A T , and then perform
triangular solves with the resulting sparse factors to obtain the step λ from (14.44a).
The steps s and x are recovered from (14.44b) and (14.44c). General-purpose sparse
Cholesky software can be applied to AD 2 A T , but modifications are needed because AD 2 A T
may be ill-conditioned or singular. (Ill conditioning of this system is often observed during
14.3.
OTHER PRIMAL-DUAL ALGORITHMS
AND
EXTENSIONS
the final stages of a primal-dual algorithm, when the elements of the diagonal weighting
matrix D 2 take on both huge and tiny values.) The Cholesky technique may encounter
diagonal elements that are very small, zero or (because of roundoff error) slightly negative.
One approach for handling this eventuality is to skip a step of the factorization, setting
the component of λ that corresponds to the faulty diagonal element to zero. We refer to
Wright [317] for details of this and other approaches.
A disadvantage of the normal-equations formulation is that if A contains any dense
columns, the entire matrix AD 2 A T is also dense. Hence, practical software identifies
dense and nearly-dense columns, excludes them from the matrix product AD 2 A T , and
performs the Cholesky factorization of the resulting sparse matrix. Then, a device such as a
Sherman-Morrison-Woodbury update is applied to account for the excluded columns. We
refer the reader to Wright [316, Chapter 11] for further details.
The formulation (14.42) has received less attention than (14.44), mainly because
algorithms and software for factoring sparse symmetric indefinite matrices are more complicated, slower, and less prevalent than sparse Cholesky algorithms. Nevertheless, the
formulation (14.42) is cleaner and more flexible than (14.44) in a number of respects.
It normally avoids the fill-in caused by dense columns in A in the matrix product AD 2 A T .
Moreover, it allows free variables (components of x with no explicit lower or upper bounds)
to be handled directly in the formulation. (The normal equations form must resort to various artificial devices to express such variables, otherwise it is not possible to perform the
block elimination that leads to the system (14.44a).)
14.3
OTHER PRIMAL-DUAL ALGORITHMS AND EXTENSIONS
OTHER PATH-FOLLOWING METHODS
Framework 14.1 is the basis of a number of other algorithms of the path-following
variety. They are less important from a practical viewpoint, but we mention them here
because of their elegance and their strong theoretical properties.
Some path-following methods choose conservative values for the centering parameter
σ (that is, σ only slightly less than 1) so that unit steps (that is, a steplength of α 1) can be
taken along the resulting direction from (14.16) without leaving the chosen neighborhood.
These methods, which are known as short-step path-following methods, make only slow
progress toward the solution because they require the iterates to stay inside a restrictive N2
neighborhood (14.17). From a theoretical point of view, however, they have the advantage
of better complexity. (A result similar to Theorem 14.4 holds with n replaced by n 1/2 in the
complexity estimate.)
Better results are obtained with the predictor-corrector method, due to Mizuno, Todd,
and Ye [208], which uses two N2 neighborhoods, nested one inside the other. (Despite the
similar terminology, this algorithm is quite distinct from Algorithm 14.3 of Section 14.2.)
Every second step of this method is a predictor step, which starts in the inner neighborhood
413
414
CHAPTER 14.
INTERIOR-POINT METHODS
and moves along the affine-scaling direction (computed by setting σ 0 in (14.16)) to
the boundary of the outer neighborhood. The gap between neighborhood boundaries is
wide enough to allow this step to make significant progress in reducing µ. Alternating with
the predictor steps are corrector steps (computed with σ 1 and α 1), which take
the next iterate back inside the inner neighborhood in preparation for the next predictor
step. The predictor-corrector algorithm produces a sequence of duality measures µk that
converge superlinearly to zero, in contrast to the linear convergence that characterizes most
methods.
POTENTIAL-REDUCTION METHODS
Potential-reduction methods take steps of the same form as path-following methods,
but they do not explicitly follow the central path C and can be motivated independently of
it. They use a logarithmic potential function to measure the worth of each point in F o and
aim to achieve a certain fixed reduction in this function at each iteration. The primal-dual
potential function, which we denote generically by , usually has two important properties:
→ ∞ if xi si → 0 for some i, while µ x T s/n → 0,
(14.45a)
→ −∞ if and only if (x, λ, s) → .
(14.45b)
The first property (14.45a) prevents any one of the pairwise products xi si from approaching
zero independently of the others, and therefore keeps the iterates away from the boundary
of the nonnegative orthant. The second property (14.45b) relates to the solution set .
If our algorithm forces to −∞, then (14.45b) ensures that the sequence approaches the
solution set.
An interesting primal-dual potential function is defined by
ρ (x, s) ρ log x T s −
n
log xi si ,
(14.46)
i1
for some parameter ρ > n (see Tanabe [283] and Todd and Ye [287]). Like all algorithms
based on Framework 14.1, potential-reduction algorithms obtain their search directions by
solving (14.10), for some σk ∈ (0, 1), and they take steps of length αk along these directions.
For instance, the step length αk may be chosen to approximately minimize ρ along the
√
computed direction. By fixing σk n/(n + n) for all k, one can guarantee constant
reduction in ρ at every iteration. Hence, ρ will approach −∞, forcing convergence.
Adaptive and heuristic choices of σk and αk are also covered by the theory, provided that
they at least match the reduction in ρ obtained from the conservative theoretical values of
these parameters.
14.3.
OTHER PRIMAL-DUAL ALGORITHMS
AND
EXTENSIONS
EXTENSIONS
Primal-dual methods for linear programming can be extended to wider classes of
problems. There are simple extensions of the algorithm to the monotone linear complementarity problem (LCP) and convex quadratic programming problems for which the
convergence and polynomial complexity properties of the linear programming algorithms
are retained. The monotone LCP is the problem of finding vectors x and s in IRn that satisfy
the following conditions:
s M x + q,
(x, s) ≥ 0,
x T s 0,
(14.47)
where M is a positive semidefinite n × n matrix and q ∈ IRn . The similarity between (14.47)
and the KKT conditions (14.3) is obvious: The last two conditions in (14.47) correspond to
(14.3d) and (14.3c), respectively, while the condition s M x + q is similar to the equations
(14.3a) and (14.3b). For practical instances of the problem (14.47), see Cottle, Pang, and
Stone [80]. Interior-point methods for monotone LCP have a close correspondence to
algorithms for linear programming. The duality measure (14.6) is redefined to be the
complementarity measure (with the same definition µ x T s/n), and the conditions that
must be satisfied by the solution can be stated similarly to (14.4) as follows:
Mx + q − s
0,
X Se
(x, s) ≥ 0.
The general formula for a path-following step is defined analogously to (14.9) as follows:
M
−I
S
X
x
s
−(M x + q − s)
−X Se + σ µe
,
where σ ∈ [0, 1]. Using these and similar adaptations, an extension of the practical method
of Section 14.2 can also be derived.
Extensions to convex quadratic programs are discussed in Section 16.6. Their
adaptation to nonlinear programming problems is the subject of Chapter 19.
Interior-point methods are highly effective in solving semidefinite programming problems, a class of problems involving symmetric matrix variables that are constrained to be
positive semidefinite. Semidefinite programming, which has been the topic of concentrated
research since the early 1990s, has applications in many areas, including control theory and
combinatorial optimization. Further information on this increasingly important topic can
be found in the survey papers of Todd [285] and Vandenberghe and Boyd [292] and the
books of Nesterov and Nemirovskii [226], Boyd et al. [37], and Boyd and Vandenberghe [38].
415
416
CHAPTER 14.
14.4
INTERIOR-POINT METHODS
PERSPECTIVES AND SOFTWARE
The appearance of interior-point methods in the 1980s presented the first serious challenge
to the dominance of the simplex method as a practical means of solving linear programming
problems. By about 1990, interior-point codes had emerged that incorporated the techniques
described in Section 14.2 and that were superior on many large problems to the simplex
codes available at that time. The years that followed saw significant improvements in simplex
software, evidenced by the appearance of packages such as CPLEX and XPRESS-MP. These
improvements were due to algorthmic advances such as steepest-edge pivoting (see Goldfarb
and Forrest [133]) and improved pricing heuristics, and also to close attention to the nuts
and bolts of efficient implementation. The efficiency of interior-point codes also continued
to improve, through improvements in the linear algebra for solving the step equations and
through the use of higher-order correctors in the step calculation (see Gondzio [138]).
During this period, a number of good interior-point codes became freely available (such
as PCx [84], HOPDM [137], BPMPD, and LIPSOL [321]) and found their way into many
applications.
In general, simplex codes are faster on problems of small-medium dimensions, while
interior-point codes are competitive and often faster on large problems. However, this
rule is certainly not hard-and-fast; it depends strongly on the structure of the particular
application. Interior-point methods are generally not able to take full advantage of prior
knowledge about the solution, such as an estimate of the solution itself or an estimate of
the optimal basis. Hence, interior-point methods are less useful than simplex approaches
in situations in which “warm-start” information is readily available. One situation of this
type involves branch-and-bound algorithms for solving integer programs, where each node
in the branch-and-bound tree requires the solution of a linear program that differs only
slightly from one already solved in the parent node. In other situations, we may wish to
solve a sequence of linear programs in which the data is perturbed slightly to investigate
sensitivity of the solutions to various perturbations, or in which we approximate a nonlinear optimization problem by a sequence of linear programs. Yıldırım and Wright [319]
describe how a given point (such as an approximate solution) can be modified to obtain
a starting point that is theoretically valid, in that it allows complexity results to be proved
that depend on the quality of the given point. In practice, however, these techniques can
be expected to provide only a modest improvement in algorithmic performance (perhaps
a factor of between 2 and 5) over a “cold” starting point such as the one described in
Section 14.2.
Interior-point software has the advantage that it is easy to program, relative to the
simplex method. The most complex operation is the solution of the large linear systems
at each iteration to compute the step; software to perform this linear algebra operation is
readily available. The interior-point code LIPSOL [321] is written entirely in the Matlab
language, apart from a small amount of FORTRAN code that interfaces to the linear algebra
software. The code PCx [84] is written in C, but also is easy for the interested user to
comprehend and modify. It is even possible for a non-expert in optimization to write an
14.4.
PERSPECTIVES
AND
SOFTWARE
efficient interior-point implementation from scratch that is customized to their particular
application.
NOTES AND REFERENCES
For more details on the material of this chapter, see the book by Wright [316].
As noted in the text, Karmarkar’s method arose from a search for linear programming
algorithms with better worst-case behavior than the simplex method. The first algorithm
with polynomial complexity, Khachiyan’s ellipsoid algorithm [180], was a computational
disappointment. In contrast, the execution times required by Karmarkar’s method were
not too much greater than simplex codes at the time of its introduction, particularly for
large linear programs. Karmarkar’s is a primal algorithm; that is, it is described, motivated, and implemented purely in terms of the primal problem (14.1) without reference
to the dual. At each iteration, Karmarkar’s algorithm performs a projective transformation on the primal feasible set that maps the current iterate x k to the center of the
set and takes a step in the feasible steepest descent direction for the transformed space.
Progress toward optimality is measured by a logarithmic potential function. Descriptions of the algorithm can be found in Karmarkar’s original paper [175] and in Fletcher
[101, Section 8.7].
Karmarkar’s method falls outside the scope of this chapter, and in any case, its practical performance does not appear to match the most efficient primal-dual methods. The
algorithms we discussed in this chapter have polynomial complexity, like Karmarkar’s
method.
Many of the algorithmic ideas that have been examined since 1984 actually had their
genesis in three works that preceded Karmarkar’s paper. The first of these is the book
of Fiacco and McCormick [98] on logarithmic barrier functions (originally proposed by
Frisch [115]), which proves existence of the central path, among many other results. Further
analysis of the central path was carried out by McLinden [205], in the context of nonlinear
complementarity problems. Finally, there is Dikin’s paper [94], in which an interior-point
method known as primal affine-scaling was originally proposed. The outburst of research on
primal-dual methods, which culminated in the efficient software packages available today,
dates to the seminal paper of Megiddo [206].
Todd gives an excellent survey of potential reduction methods in [284]. He relates
the primal-dual potential reduction method mentioned above to pure primal potential
reduction methods, including Karmarkar’s original algorithm, and discusses extensions to
special classes of nonlinear problems.
For an introduction to complexity theory and its relationship to optimization, see the
book by Vavasis [297].
Andersen et al. [6] cover many of the practical issues relating to implementation of
interior-point methods. In particular, they describe an alternative scheme for choosing the
initial point, for the case in which upper bounds are also present on the variables.
417
418
CHAPTER 14.
✐
INTERIOR-POINT METHODS
EXERCISES
✐
14.1 This exercise illustrates the fact that the bounds (x, s) ≥ 0 are essential in
relating solutions of the system (14.4a) to solutions of the linear program (14.1) and its
dual. Consider the following linear program in IR2 :
min x1 , subject to x1 + x2 1,
(x1 , x2 ) ≥ 0.
Show that the primal-dual solution is
x∗ 0
,
1
λ∗ 0,
s∗ 1
.
0
Also verify that the system F(x, λ, s) 0 has the spurious solution
x
1
0
,
λ 1,
s
0
−1
,
which has no relation to the solution of the linear program.
✐
14.2
(i) Show that N2 (θ1 ) ⊂ N2 (θ2 ) when 0 ≤ θ1 < θ2 < 1 and that N−∞ (γ1 ) ⊂ N−∞ (γ2 )
for 0 < γ2 ≤ γ1 ≤ 1.
(ii) Show that N2 (θ ) ⊂ N−∞ (γ ) if γ ≤ 1 − θ .
✐
14.3 Given an arbitrary point (x, λ, s) ∈ F o , find the range of γ values for which
(x, λ, s) ∈ N−∞ (γ ). (The range depends on x and s.)
✐
14.4 For n 2, find a point (x, s) > 0 for which the condition
X Se − µe2 ≤ θ µ
is not satisfied for any θ ∈ [0, 1).
✐
14.5 Prove that the neighborhoods N−∞ (1) (see (14.18)) and N2 (0) (see (14.17))
coincide with the central path C.
✐
14.6 In the long-step path-following method (Algorithm 14.2), give a procedure for
calculating the maximum value of α such that (14.20) is satisfied.
✐
✐
14.7 Show that ρ defined by (14.46) has the property (14.45a).
14.8 Prove that the coefficient matrix in (14.16) is nonsingular if and only if A has
full row rank.
14.4.
✐
PERSPECTIVES
AND
SOFTWARE
14.9 Given (x, λ, s) satisfying (14.10), prove (14.23).
✐
pri
14.10 Given an iterate (x k , λk , s k ) with (x k , s k ) > 0, show that the quantities αmax
dual
and αmax
defined by (14.36) are the largest values of α such that x k + αx k ≥ 0 and
s k + αs k ≥ 0, respectively.
✐
14.11 Verify (14.37).
✐ 14.12 Given that X and S are diagonal with positive diagonal elements, show that
the coefficient matrix in (14.44a) is symmetric and positive definite if and only if A has full
row rank. Does this result continue to hold if we replace D by a diagonal matrix in which
exactly m of the diagonal elements are positive and the remainder are zero? (Here m is the
number of rows of A.)
✐
14.13 Given a point (x, λ, s) with (x, s) > 0, consider the trajectory H defined by
⎡
⎢
F x̂(τ ), λ̂(τ ), ŝ(τ ) ⎣
(1 − τ )(A T λ + s − c)
(1 − τ )(Ax − b)
⎤
⎥
⎦,
(x̂(τ ), ŝ(τ )) ≥ 0,
(1 − τ )X Se
for τ ∈ [0, 1], and note that x̂(0), λ̂(0), ŝ(0) (x, λ, s), while the limit of
x̂(τ ), λ̂(τ ), ŝ(τ ) as τ ↑ 1 will lie in the primal-dual solution set of the linear program. Find equations for the first, second, and third derivatives of H with respect to
τ at τ 0. Hence, write down a Taylor series approximation to H near the point
(x, λ, s).
✐
14.14 Consider the following linear program, which contains “free variables” denoted
by y:
min c T x + d T y, subject to A1 x + A2 y b, x ≥ 0.
By introducing Lagrange multipliers λ for the equality constraints and s for the bounds
x ≥ 0, write down optimality conditions for this problem in an analogous fashion to (14.3).
Following (14.4) and (14.16), use these conditions to derive the general step equations for
a primal-dual interior-point method. Express these equations in augmented system form
analogously to (14.42) and explain why it is not possible to reduce further to a formulation
like (14.44) in which the coefficient matrix is symmetric positive definite.
✐
14.15 Program Algorithm 14.3 in Matlab. Choose η 0.99 uniformly in (14.38).
Test your code on a linear programming problem (14.1) generated by choosing A randomly,
419
420
CHAPTER 14.
INTERIOR-POINT METHODS
and then setting x, s, b, and c as follows:
xi si random positive number
i 1, 2, . . . , m,
0
i m + 1, m + 2, . . . , n,
random positive number
i m + 1, m + 2, . . . , n
0
i 1, 2, . . . , m,
λ random vector,
c A T λ + s,
b Ax.
Choose the starting point (x 0 , λ0 , s 0 ) with the components of x 0 and s 0 set to large positive
values.
✐
14.17 Show that the solutions of the problems (14.39) are given explicitly by (14.40).
This is page 421
Printer: Opaque this
15
CHAPTER
Fundamentals
of Algorithms
for Nonlinear
Constrained
Optimization
In this chapter, we begin our discussion of algorithms for solving the general constrained
optimization problem
minn f (x)
x∈IR
subject to
ci (x) 0,
i ∈ E,
ci (x) ≥ 0,
i ∈ I,
(15.1)
where the objective function f and the constraint functions ci are all smooth, real-valued
functions on a subset of IRn , and I and E are finite index sets of inequality and equality
constraints, respectively. In Chapter 12, we used this general statement of the problem
422
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
to derive optimality conditions that characterize its solutions. This theory is useful for
motivating the various algorithms discussed in the remainder of the book, which differ from
each other in fundamental ways but are all iterative in nature. They generate a sequence of
estimates of the solution x ∗ that, we hope, tend toward a solution. In some cases, they also
generate a sequence of guesses for the Lagrange multipliers associated with the constraints.
As in the chapters on unconstrained optimization, we study only algorithms for finding local
solutions of (15.1); the problem of finding a global solution is outside the scope of this book.
We note that this chapter is not concerned with individual algorithms themselves,
but rather with fundamental concepts and building blocks that are common to more than
one algorithm. After reading Sections 15.1 and 15.2, the reader may wish to glance at the
material in Sections 15.3, 15.4, 15.5, and 15.6, and return to these sections as needed during
study of subsequent chapters.
15.1
CATEGORIZING OPTIMIZATION ALGORITHMS
We now catalog the algorithmic approaches presented in the rest of the book. No standard
taxonomy exists for nonlinear optimization algorithms; in the remaining chapters we have
grouped the various approaches as follows.
I. In Chapter 16 we study algorithms for solving quadratic programming problems. We
consider this category separately because of its intrinsic importance, because its particular
characteristics can be exploited by efficient algorithms, and because quadratic programming
subproblems need to be solved by sequential quadratic programming methods and certain
interior-point methods for nonlinear programming. We discuss active set, interior-point,
and gradient projection methods.
II. In Chapter 17 we discuss penalty and augmented Lagrangian methods. By combining the
objective function and constraints into a penalty function, we can attack problem (15.1) by
solving a sequence of unconstrained problems. For example, if only equality constraints are
present in (15.1), we can define the quadratic penalty function as
f (x) +
µ 2
c (x),
2 i∈E i
(15.2)
where µ > 0 is referred to as a penalty parameter. We minimize this unconstrained function,
for a series of increasing values of µ, until the solution of the constrained optimization
problem is identified to sufficient accuracy.
If we use an exact penalty function, it may be possible to find a local solution of (15.1) by
solving a single unconstrained optimization problem. For the equality-constrained problem,
15.1.
CATEGORIZING OPTIMIZATION ALGORITHMS
the function defined by
f (x) + µ
|ci (x)|,
i∈E
is usually an exact penalty function, for a sufficiently large value of µ > 0. Although they
often are nondifferentiable, exact penalty functions can be minimized by solving a sequence
of smooth subproblems.
In augmented Lagrangian methods, we define a function that combines the properties
of the Lagrangian function (12.33) and the quadratic penalty function (15.2). This so-called
augmented Lagrangian function has the following form for equality-constrained problems:
L A (x, λ; µ) f (x) −
i∈E
λi ci (x) +
µ 2
c (x).
2 i∈E i
Methods based on this function fix λ to some estimate of the optimal Lagrange multiplier
vector and fix µ to some positive value, then find a value of x that approximately minimizes
L A (·, λ; µ). At this new x-iterate, λ and µ may be updated; then the process is repeated.
This approach avoids certain drawbacks associated with the minimization of the quadratic
penalty function (15.2).
III. In Chapter 18 we describe sequential quadratic programming (SQP) methods, which
model (15.1) by a quadratic programming subproblem at each iterate and define the search
direction to be the solution of this subproblem. In the basic SQP method, we define the
search direction pk at the iterate (xk , λk ) to be the solution of
min
p
subject to
1
2
p T ∇x2x L(xk , λk ) p + ∇ f (xk )T p
(15.3a)
∇ci (xk )T p + ci (xk ) 0,
i ∈ E,
(15.3b)
∇ci (xk ) p + ci (xk ) ≥ 0,
i ∈ I,
(15.3c)
T
where L is the Lagrangian function defined in (12.33). The objective in this subproblem is
an approximation to the change in the Lagrangian function in moving from xk to xk + p,
while the constraints are linearizations of the constraints in (15.1). A trust-region constraint
may be added to (15.3) to control the length and quality of the step, and quasi-Newton
approximate Hessians can be used in place of ∇x2x L(xk , λk ). In a variant called sequential
linear-quadratic programming, the step pk is computed in two stages. First, we solve a linear
program that is defined by omitting the first (quadratic) term from the objective (15.3a)
and adding a trust-region constraint to (15.3). Next, we obtain the step pk by solving an
equality-constrained subproblem in which the constraints active at the solution of the linear
program are imposed as equalities, while all other constraints are ignored.
IV. In Chapter 19 we study interior-point methods for nonlinear programming. These methods can be viewed as extensions of the primal-dual interior-point methods for linear
423
424
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
programming discussed in Chapter 14. We can also view them as barrier methods that
generate steps by solving the problem
min
x,s
subject to
f (x) − µ
m
log si
(15.4a)
i ∈ E,
(15.4b)
i1
ci (x) 0,
ci (x) − si 0,
i ∈ I,
(15.4c)
for some positive value of the barrier parameter µ, where the variables si > 0 are slacks.
Interior-point methods constitute the newest class of methods for nonlinear programming
and have already proved to be formidable competitors of sequential quadratic programming
methods.
The algorithms in categories I, III, and IV make use of elimination techniques, in which
the constraints are used to eliminate some of the degrees of freedom in the problem. As a
background to those algorithms, we discuss elimination in Section 15.3. In later sections
we discuss merit functions and filters, which are important mechanisms for promoting
convergence of nonlinear programming algorithms from remote starting points.
15.2
THE COMBINATORIAL DIFFICULTY OF
INEQUALITY-CONSTRAINED PROBLEMS
One of the main challenges in solving nonlinear programming problems lies in dealing with
inequality constraints—in particular, in deciding which of these constraints are active at
the solution and which are not. One approach, which is the essence of active-set methods,
starts by making a guess of the optimal active set A∗ , that is, the set of constraints that
are satisfied as equalities at a solution. We call our guess the working set and denote it by
W. We then solve a problem in which the constraints in the working set are imposed as
equalities and the constraints not in W are ignored. We then check to see if there is a choice
of Lagrange multipliers such that the solution x ∗ obtained for this W satisfies the KKT
conditions (12.34). If so, we accept x ∗ as a local solution of (15.1). Otherwise, we make a
different choice of W and repeat the process. This approach is based on the observation
that, in general, it is much simpler to solve equality-constrained problems than to solve
nonlinear programs.
The number of choices for working set W may be very large—up to 2|I| , where |I|
is the number of inequality constraints. We arrive at this estimate by observing that we can
make one of two choices for each i ∈ I: to include it in W or leave it out. Since the number of
possible working sets grows exponentially with the number of inequalities—a phenomenon
which we refer to as the combinatorial difficulty of nonlinear programming—we cannot
hope to design a practical algorithm by considering all possible choices for W.
15.2.
THE COMBINATORIAL DIFFICULTY
OF
INEQUALITY-CONSTRAINED PROBLEMS
The following example suggests that even for a small number of inequality constraints,
determination of the optimal active set is not a simple task.
❏ EXAMPLE 15.1
Consider the problem
def
min f (x, y) 12 (x − 2)2 + 12 (y − 12 )2
x,y
(x + 1)−1 − y −
1
4
(15.5)
≥ 0,
x ≥ 0,
subject to
y ≥ 0.
We label the constraints, in order, with the indices 1 through 3. Figure 15.1 illustrates the
contours of the objective function (dashed circles). The feasible region is the region enclosed
by the curve and the two axes. We see that only the first constraint is active at the solution,
which is (x ∗ , y ∗ )T (1.953, 0.089)T .
Let us now apply the working-set approach described above to (15.5), considering all
3
2 8 possible choices of W.
We consider first the possibility that no constraints are active at the solution, that is,
W ∅. Since ∇ f (x − 2, y − 1/2)T , we see that the unconstrained minimum of f lies
outside the feasible region. Hence, the optimal active set cannot be empty.
There are seven further possibilities. First, all three constraints could be active (that is,
W {1, 2, 3}). A glance at Figure 15.1 shows that this does not happen for our problem; the
three constraints do not share a common point of intersection. Three further possibilities are
obtained by making a single constraint active (that is, W {1}, W {2}, and W {3}),
y
(2,0.5)
(x*,y*)
x
Figure 15.1 Graphical illustration of problem (15.5).
425
426
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
while the final three possibilities are obtained by making exactly two constraints active (that
is, W {1, 2}, W {1, 3}, and W {2, 3}). We consider three of these cases in detail.
- W {2}; that is, only the constraint x 0 is active. If we minimize f enforcing only
this constraint, we obtain the point (0, 1/2)T . A check of the KKT conditions (12.34)
shows that no matter how we choose the Lagrange multipliers, we cannot satisfy all
these conditions at (0, 1/2)T . (We must have λ1 λ3 0 to satisfy (12.34e), which
implies that we must set λ2 −2 to satisfy (12.34a); but this value of λ2 violates the
condition (12.34d).)
- W {1, 3}, which yields the single feasible point (3, 0)T . Since constraint 2 is
inactive at this point, we have λ2 0, so by solving (12.34a) for the other Lagrange
multipliers, we obtain λ1 −16 and λ3 −16.5. These values are negative, so they
violate (12.34d), and x (3, 0)T cannot be a solution of (15.1).
- W {1}. Solving the equality-constrained problem in which the first constraint is
active, we obtain (x, y)T (1.953, 0.089)T with Lagrange multiplier λ1 0.411. It
is easy to see that by setting λ2 λ3 0, the remaining KKT conditions (12.34) are
satisfied, so we conclude that this is a KKT point. Furthermore, it is easy to show that
the second-order sufficient conditions are satisfied, as the Hessian of the Lagrangian
is positive definite.
❐
Even for this small example, we see that it is exhausting to consider all possible choices
for W. Figure 15.1 suggests, however, that some choices of W can be eliminated from
consideration if we make use of knowledge of the functions that define the problem, and
their derivatives. In fact, the active set methods described in Chapter 16 use this kind of
information to make a series of educated guesses for the working set, avoiding choices of W
that obviously will not lead to a solution of (15.1).
A different approach is followed by interior-point (or barrier) methods discussed in
Chapter 19. These methods generate iterates that stay away from the boundary of the feasible
region defined by the inequality constraints. As the solution of the nonlinear program is
approached, the barrier effects are weakened to permit an increasingly accurate estimate of
the solution. In this manner, interior-point methods avoid the combinatorial difficulty of
nonlinear programming.
15.3
ELIMINATION OF VARIABLES
When dealing with constrained optimization problems, it is natural to try to use the constraints to eliminate some of the variables from the problem, to obtain a simpler problem
with fewer degrees of freedom. Elimination techniques must be used with care, however, as
they may alter the problem or introduce ill conditioning.
15.3.
ELIMINATION
OF
VARIABLES
y 2 = (x-1) 3
y
x 2+y 2= 4
x 2+y 2=1
x
(1,0)
Figure 15.2
The danger of nonlinear
elimination.
We begin with an example in which it is safe and convenient to eliminate variables. In
the problem
min f (x) f (x1 , x2 , x3 , x4 )
subject to
x1 + x32 − x4 x3 0,
−x2 + x4 + x32 0,
there is no risk in setting
x1 x4 x3 − x32 ,
x2 x4 + x32 ,
to obtain a function of two variables
h(x3 , x4 ) f (x4 x3 − x32 , x4 + x32 , x3 , x4 ),
which we can minimize using the unconstrained optimization techniques described in
earlier chapters.
The dangers of nonlinear elimination are illustrated in the following example.
❏ EXAMPLE 15.2
(FLETCHER [101])
Consider the problem
min x 2 + y 2
subject to (x − 1)3 y 2 .
The contours of the objective function and the constraints are illustrated in Figure 15.2,
which shows that the solution is (x, y) (1, 0).
427
428
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
We attempt to solve this problem by eliminating y. By doing so, we obtain
h(x) x 2 + (x − 1)3 .
Clearly, h(x) → −∞ as x → −∞. By blindly applying this transformation we may
conclude that the problem is unbounded, but this view ignores the fact that the constraint (x − 1)3 y 2 implicitly imposes the bound x ≥ 1 that is active at the solution.
Hence, if we wish to eliminate y, we should explicitly introduce the bound x ≥ 1 into the
problem.
❐
This example shows that the use of nonlinear equations to eliminate variables may
result in errors that can be difficult to trace. For this reason, nonlinear elimination is not
used by most optimization algorithms. Instead, many algorithms linearize the constraints
and apply elimination techniques to the simplified problem. We now describe systematic
procedures for performing variable elimination using linear constraints.
SIMPLE ELIMINATION USING LINEAR CONSTRAINTS
We consider the minimization of a nonlinear function subject to a set of linear equality
constraints,
subject to Ax b,
min f (x)
(15.6)
where A is an m × n matrix with m ≤ n. Suppose for simplicity that A has full row rank.
(If such is not the case, we find either that the problem is inconsistent or that some of the
constraints are redundant and can be deleted without affecting the solution of the problem.)
Under this assumption, we can find a subset of m columns of A that is linearly independent.
If we gather these columns into an m × m matrix B and define an n × n permutation matrix
P that swaps these columns to the first m column positions in A, we can write
A P [B | N ],
(15.7)
where N denotes the n − m remaining columns of A. (The notation here is consistent
with that of Chapter 13, where we discussed similar concepts in the context of linear
programming.) We define the subvectors xB ∈ IRm and xN ∈ IRn−m as follows:
xB
xN
P T x,
(15.8)
15.3.
ELIMINATION
OF
VARIABLES
and call xB the basic variables and B the basis matrix. Noting that P P T I , we can rewrite
the constraint Ax b as
b Ax A P(P T x) BxB + N xN .
By rearranging this formula, we deduce that the basic variables can be expressed as follows:
xB B −1 b − B −1 N xN .
(15.9)
We can therefore compute a feasible point for the constraints Ax b by choosing any value
of xN and then setting xB according to the formula (15.9). The problem (15.6) is therefore
equivalent to the unconstrained problem
def
min h(xN ) f
xN
P
B −1 b − B −1 N xN
.
xN
(15.10)
We refer to the substitution in (15.9) as simple elimination of variables.
This discussion shows that a nonlinear optimization problem with linear equality
constraints is, from a mathematical point of view, the same as an unconstrained problem.
❏ EXAMPLE 15.3
Consider the problem
1
min sin(x1 + x2 ) + x32 + (x4 + x54 + x6 /2)
3
8x1 − 6x2 + x3 + 9x4 + 4x5 6
subject to
3x1 + 2x2 − x4 + 6x5 + 4x6 −4.
(15.11a)
(15.11b)
By defining the permutation matrix P so as to reorder the components of x as x T (x3 , x6 , x1 , x2 , x4 , x5 )T , we find that the coefficient matrix A P is
AP 1
0
8
−6
9
4
0
4
3
2
−1
6
.
The basis matrix B is diagonal and therefore easy to invert. We obtain from (15.9) that
x3
x6
⎡
8
−⎣ 3
4
−6
1
2
9
1
−
4
⎤
⎡
x1
⎢
4
⎢ x2
3 ⎦⎢
⎢ x
⎣ 4
2
x5
⎤
⎥ ⎥
6
⎥+
.
⎥
−1
⎦
(15.12)
429
430
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
By substituting for x3 and x6 in (15.11a), the problem becomes
min sin(x1 + x2 ) + (8x1 − 6x2 + 9x4 + 4x5 − 6)2
(15.13)
x1 ,x2 ,x4 ,x5
1
+ (x4 + x54 − [(1/2) + (3/8)x1 + (1/4)x2 − (1/8)x4 + (3/4)x5 ]).
3
We could have chosen two other columns of the coefficient matrix A (that is, two
variables other than x3 and x6 ) as the basis for elimination in the system (15.11b), but the
matrix B −1 N would not have been so simple.
❐
A set of m independent columns can be selected, in general, by means of Gaussian
elimination. In the parlance of linear algebra, we can compute the row echelon form of the
matrix and choose the pivot columns as the columns of the basis B. Ideally, we would like
B to be easy to factor and well conditioned. A technique that suits these purposes is a sparse
Gaussian elimination approach that attempts to preserve sparsity while keeping rounding
errors under control. A well-known implementation of this algorithm is MA48 from the
HSL library [96]. As we discuss below, however, there is no guarantee that the Gaussian
elimination process will identify the best choice of basis matrix.
There is an interesting interpretation of the simple elimination-of-variables approach
that we have just described. To simplify the notation, we will assume from now on that
the coefficient matrix is already given to us so that the basic columns appear in the first m
positions, that is, P I .
From (15.8) and (15.9) we see that any feasible point x for the linear constraints in
(15.6) can be written as
xB
x Y b + Z xN ,
xN
(15.14)
where
Y B −1
0
,
Z
−B −1 N
I
.
(15.15)
Note that Z has n − m linearly independent columns (because of the presence of the identity
matrix in the lower block) and that it satisfies AZ 0. Therefore, Z is a basis for the null
space of A. In addition, the columns of Y and the columns of Z form a linearly independent
set. We note also from (15.15), (15.7) that Y b is a particular solution of the linear constraints
Ax b.
In other words, the simple elimination technique expresses feasible points as the sum
of a particular solution of Ax b (the first term in (15.14)) plus a displacement along the
15.3.
ELIMINATION
OF
VARIABLES
x2
Ax = b
x1
coordinate relaxation step
Figure 15.3 Simple elimination, showing the coordinate relaxation step obtained by
choosing the basis to be the first column of A.
null space of the constraints (the second term in (15.14)). The relations (15.14), (15.15)
indicate that the particular Y b solution is obtained by holding n −m components of x at zero
while relaxing the other m components (the ones in xB ) until they reach the constraints. The
particular solution Y b is sometimes known as the coordinate relaxation step. In Figure 15.3,
we see the coordinate relaxation step Y b obtained by choosing the basis matrix B to be the
first column of A. If we were to choose B to be the second column of A, the coordinate
relaxation step would lie along the x2 axis.
Simple elimination is inexpensive but can give rise to numerical instabilities. If the
feasible set in Figure 15.3 consisted of a line that was almost parallel to the x1 axis, the
coordinate relaxation along this axis would be very large in magnitude. We would then be
computing x as the difference of very large vectors, giving rise to numerical cancellation.
In that situation it would be preferable to choose a particular solution along the x2 axis,
that is, to select a different basis. Selection of the best basis is, therefore, not a straightforward task in general. To overcome the dangers of an excessively large coordinate relaxation
step, we could define the particular solution Y b as the minimum-norm step to the constraints. This approach is a special case of more general elimination strategies, which we now
describe.
GENERAL REDUCTION STRATEGIES FOR LINEAR CONSTRAINTS
To generalize (15.14) and (15.15), we choose matrices Y ∈ IRn×m and Z ∈ IRn×(n−m)
with the following properties:
[Y | Z ] ∈ IRn×n is nonsingular,
AZ 0.
(15.16)
431
432
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
These properties indicate that, as in (15.15), the columns of Z are a basis for the null space
of A. Since A has full row rank, so does A[Y | Z ] [AY | 0], so it follows that the m × m
matrix AY is nonsingular. We now express any solution of the linear constraints Ax b as
x Y xY + Z xZ ,
(15.17)
for some vectors xY ∈ IRm and xZ ∈ IRn−m . By substituting (15.17) into the constraints
Ax b, we obtain
Ax (AY )xY b;
hence by nonsingularity of AY , xY can be written explicitly as
xY (AY )−1 b.
(15.18)
By substituting this expression into (15.17), we conclude that any vector x of the form
x Y (AY )−1 b + Z xZ
(15.19)
satisfies the constraints Ax b for any choice of xZ ∈ IRn−m . Therefore, the problem (15.6)
can be restated equivalently as the following unconstrained problem
min f (Y (AY )−1 b + Z xZ ).
xZ
(15.20)
Ideally, we would like to choose Y in such a way that the matrix AY is as well
conditioned as possible, since it needs to be factorized to give the particular solution
Y (AY )−1 b. We can do this by computing Y and Z by means of a QR factorization of
A T , which has the form
R
,
(15.21)
AT Q 1 Q 2
0
where Q 1 Q 2 is orthogonal. The submatrices Q 1 and Q 2 have orthonormal columns
and are of dimension n × m and n × (n − m), while R is m × m upper triangular and
nonsingular and is an m × m permutation matrix. (See the discussion following (A.24)
in the Appendix for further details.) We now define
Y Q1,
Z Q2,
(15.22)
so that the columns of Y and Z form an orthonormal basis of IRn . If we expand (15.21) and
do a little rearrangement, we obtain
AY R T ,
AZ 0.
15.3.
ELIMINATION
OF
VARIABLES
Therefore, Y and Z have the desired properties, and the condition number of AY is the
same as that of R, which in turn is the same as that of A itself. From (15.19) we see that any
solution of Ax b can be expressed as
x Q 1 R −T T b + Q 2 xZ ,
for some vector xZ . The computation R −T T b can be carried out inexpensively, at the cost
of a single triangular substitution.
A simple computation shows that the particular solution Q 1 R −T T b can also be
written as A T (A A T )−1 b. This vector is the solution of the following problem:
min x2
subject to Ax b;
that is, it is the minimum-norm solution of Ax b. See Figure 15.5 for an illustration of
this step.
Elimination via the orthogonal basis (15.22) is ideal from the point of view of numerical stability. The main cost associated with this reduction strategy is in computing the QR
factorization (15.21). Unfortunately, for problems in which A is large and sparse, a sparse
QR factorization can be much more costly to compute than the sparse Gaussian elimination strategy used in simple elimination. Therefore, other elimination strategies have been
developed that seek a compromise between these two techniques; see Exercise 15.7.
x2
YxY
x1
ZxZ
Ax= b
x3
Figure 15.4 General elimination: Case in which A ∈ IR1×3 , showing the particular
solution and a step in the null space of A.
433
434
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
x2
T
T -1
A [AA ] b
Ax=b
x1
Figure 15.5
The minimum-norm step.
EFFECT OF INEQUALITY CONSTRAINTS
Elimination of variables is not always beneficial if inequality constraints are present
alongside the equalities. For instance, if problem (15.11) had the additional constraint
x ≥ 0, then after eliminating the variables x3 and x6 , we would be left with the problem of
minimizing the function in (15.13) subject to the constraints
(x1 , x2 , x4 , x5 ) ≥ 0,
8x1 − 6x2 + 9x4 + 4x5 ≤ 6,
(3/4)x1 + (1/2)x2 − (1/4)x4 + (3/2)x5 ≤ −1.
Hence, the cost of eliminating the equality constraints (15.11b) is to make the inequalities
more complicated than the simple bounds x ≥ 0. For many algorithms, this transformation
will not yield any benefit.
If, however, problem (15.11) included the general inequality constraint 3x1 +2x3 ≥ 1,
the elimination (15.12) would transform the problem into one of minimizing the function
in (15.13) subject to the inequality constraint
− 13x1 + 12x2 − 18x4 − 8x5 ≥ −11.
(15.23)
In this case, the inequality constraint would not become much more complicated after elimination of the equality constraints, so it is probably worthwhile to perform the
elimination.
15.4.
15.4
MERIT FUNCTIONS
AND
FILTERS
MERIT FUNCTIONS AND FILTERS
Suppose that an algorithm for solving the nonlinear programming problem (15.1) generates
a step that reduces the objective function but increases the violation of the constraints. Should
we accept this step?
This question is not easy to answer. We must look for a way to balance the twin (often
competing) goals of reducing the objective function and satisfying the constraints. Merit
functions and filters are two approaches for achieving this balance. In a typical constrained
optimization algorithm, a step p will be accepted only if it leads to a sufficient reduction in
the merit function φ or if it is acceptable to the filter. These concepts are explained in the
rest of the section.
MERIT FUNCTIONS
In unconstrained optimization, the objective function f is the natural choice for the
merit function. All the unconstrained optimization methods described in this book require
that f be decreased at each step (or at least within a certain number of iterations). In feasible
methods for constrained optimization in which the starting point and all subsequent iterates
satisfy all the constraints in the problem, the objective function is still an appropriate merit
function. On the other hand, algorithms that allow iterates to violate the constraints require
some means to assess the quality of the steps and iterates. The merit function in this case
combines the objective with measures of constraint violation.
A popular choice of merit function for the nonlinear programming problem (15.1) is
the 1 penalty function defined by
φ1 (x; µ) f (x) + µ
i∈E
|ci (x)| + µ
[ci (x)]− ,
(15.24)
i∈I
where we use the notation [z]− max{0, −z}. The positive scalar µ is the penalty parameter,
which determines the weight that we assign to constraint satisfaction relative to minimization
of the objective. The 1 merit function φ1 is not differentiable because of the presence of the
absolute value and [·]− functions, but it has the important property of being exact.
Definition 15.1 (Exact Merit Function).
A merit function φ(x; µ) is exact if there is a positive scalar µ∗ such that for any µ > µ∗ ,
any local solution of the nonlinear programming problem (15.1) is a local minimizer of φ(x; µ).
We show in Theorem 17.3 that, under certain assumptions, the 1 merit function
φ1 (x; µ) is exact and that the threshold value µ∗ is given by
µ∗ max{|λi∗ |, i ∈ E ∪ I},
435
436
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
where the λi∗ denote the Lagrange multipliers associated with an optimal solution x ∗ . Since
the optimal Lagrange multipliers are, however, not known in advance, algorithms based on
the 1 merit function contain rules for adjusting the penalty parameter whenever there is
reason to believe that it is not large enough (or is excessively large). These rules depend on
the choice of optimization algorithm and are discussed in the next chapters.
Another useful merit function is the exact 2 function, which for equality-constrained
problems takes the form
φ2 (x; µ) f (x) + µc(x)2 .
(15.25)
This function is nondifferentiable because the 2-norm term is not squared; its derivative is
not defined at x for which c(x) 0.
Some merit functions are both smooth and exact. To ensure that both properties hold,
we must include additional terms in the merit function. For equality-constrained problems,
Fletcher’s augmented Lagrangian is given by
φF (x; µ) f (x) − λ(x)T c(x) + 12 µ
ci (x)2 ,
(15.26)
i∈E
where µ > 0 is the penalty parameter and
λ(x) [A(x)A(x)T ]−1 A(x)∇ f (x).
(15.27)
(Here A(x) denotes the Jacobian of c(x).) Although this merit function has some interesting
theoretical properties, it has practical limitations, including the expense of solving for λ(x)
in (15.27).
A quite different merit function is the (standard) augmented Lagrangian in x and λ,
which for equality-constrained problems has the form
L A (x, λ; µ) f (x) − λT c(x) + 12 µc(x)22 .
(15.28)
We assess the acceptability of a trial point (x + , λ+ ) by comparing the value of L A (x + , λ+ ; µ)
with the value at the current iterate, (x, λ). Strictly speaking, L A is not a merit function in
the sense that a solution (x ∗ , λ∗ ) of the nonlinear programming problem is not in general a
minimizer of L A (x, λ; µ) but only a stationary point. Although some sequential quadratic
programming methods use L A successfully as a merit function by adaptively modifying
µ and λ, we will not consider its use as a merit function further. Instead, we will focus
primarily on the nonsmooth exact penalty functions φ1 and φ2 .
A trial step x + x + αp generated by a line search algorithm will be accepted if it
produces a sufficient decrease in the merit function φ(x; µ). One way to define this concept
is analogous to the condition (3.4) used in unconstrained optimization, where the amount
15.4.
MERIT FUNCTIONS
AND
FILTERS
of decrease is not too small relative to the predicted change in the function over the step.
The 1 and 2 merit functions are not differentiable, but they have a directional derivative.
(See (A.51) for background on directional derivatives.) We write the directional derivative
of φ(x; µ) in the direction p as
D(φ(x; µ); p).
In a line search method, the sufficient decrease condition requires the steplength parameter
α > 0 to be small enough that the inequality
φ(x + αp; µ) ≤ φ(x; µ) + ηα D(φ(x; µ); p),
(15.29)
is satisfied for some η ∈ (0, 1).
Trust-region methods typically use a quadratic model q( p) to estimate the value of
the merit function φ after a step p; see Section 18.5. The sufficient decrease condition can
be stated in terms of a decrease in this model, as follows
φ(x + p; µ) ≤ φ(x; µ) − η(q(0) − q( p)),
(15.30)
for some η ∈ (0, 1). (The final term in (15.30) is positive, because the step p is computed
to decrease the model q.)
FILTERS
Filter techniques are step acceptance mechanisms based on ideas from multiobjective
optimization. Our derivation starts with the observation that nonlinear programming has
two goals: minimization of the objective function and the satisfaction of the constraints. If
we define a measure of infeasibility as
h(x) |ci (x)| +
i∈E
[ci (x)]− ,
(15.31)
i∈I
we can write these two goals as
min f (x)
x
and
min h(x).
x
(15.32)
Unlike merit functions, which combine both problems into a single minimization problem, filter methods keep the two goals in (15.32) separate. Filter methods accept a trial
step x + as a new iterate if the pair ( f (x + ), h(x + )) is not dominated by a previous
pair ( fl , h l ) ( f (xl ), h(xl )) generated by the algorithm. These concepts are defined as
follows.
437
438
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
h(x)
(fk ,hk )
(f i ,h i )
f(x)
Figure 15.6
Graphical illustration of a filter with four pairs.
Definition 15.2.
(a) A pair ( f k , h k ) is said to dominate another pair ( fl , h l ) if both f k ≤ fl and h k ≤ h l .
(b) A filter is a list of pairs ( fl , h l ) such that no pair dominates any other.
(c) An iterate xk is said to be acceptable to the filter if ( f k , h k ) is not dominated by any pair
in the filter.
When an iterate xk is acceptable to the filter, we (normally) add ( f k , h k ) to the filter
and remove any pairs that are dominated by ( f k , h k ). Figure 15.6 shows a filter where each
pair ( fl , h l ) in the filter is represented as a black dot. Every point in the filter creates an
(infinite) rectangular region, and their union defines the set of pairs not acceptable to the
filter. More specifically, a trial point x + is acceptable to the filter if ( f + , h + ) lies below or to
the left of the solid line in Figure 15.6.
To compare the filter and merit function approaches, we plot in Figure 15.7 the
contour line of the set of pairs ( f, h) such that f + µh f k + µh k , where xk is the current
iterate. The region to the left of this line corresponds to the set of pairs that reduce the merit
function φ(x; µ) f (x) + µh(x); clearly this set is quite different from the set of points
acceptable to the filter.
If a trial step x + xk + αk pk generated by a line search method gives a pair ( f + , h + )
that is acceptable to the filter, we set xk+1 x + ; otherwise, a backtracking line search is
performed. In a trust-region method, if the step is not acceptable to the filter, the trust
region is reduced, and a new step is computed.
Several enhancements to this filter technique are needed to obtain global convergence
and good practical performance. We need to ensure, first of all, that we do not accept a point
whose ( f, h) pair is very close to the current pair ( f k , h k ) or to another pair in the filter. We
15.4.
MERIT FUNCTIONS
AND
FILTERS
h(x)
(fk ,hk )
isovalue of
merit function
(f i ,h i )
f(x)
Figure 15.7 Comparing the filter and merit function techniques.
do so by modifying the acceptability criterion and imposing a sufficient decrease condition.
A trial iterate x + is acceptable to the filter if, for all pairs ( f j , h j ) in the filter, we have that
f (x + ) ≤ f j − βh j
or
h(x + ) ≤ h j − βh j ,
(15.33)
for β ∈ (0, 1). Although this condition is effective in practice using, say β 10−5 , for
purposes of analysis it may be advantageous to replace the first inequality by
f (x + ) ≤ f j − βh + .
A second enhancement addresses some problematic aspects of the filter mechanism.
Under certain circumstances, the search directions generated by line search methods may
require arbitrarily small steplengths αk to be acceptable to the filter. This phenomenon can
cause the algorithm to stall and fail. To guard against this situation, if the backtracking
line search generates a steplength that is smaller than a given threshold αmin , the algorithm
switches to a feasibility restoration phase, which we describe below. Similarly, in a trust-region
method, if a sequence of trial steps is rejected by the filter, the trust-region radius may be
decreased so much that the trust-region subproblem becomes infeasible (see Section 18.5).
In this case, too, the feasibility restoration phase is invoked. (Other mechanisms could be
employed to handle this situation, but as we discuss below, the feasibility restoration phase
can help the algorithm achieve other useful goals.)
The feasibility restoration phase aims exclusively to reduce the constraint violation,
that is, to find an approximate solution to the problem
min h(x).
x
439
440
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
Although h(x) defined by (15.31) is not smooth, we show in Chapter 17 how to minimize
it using a smooth constrained optimization subproblem. This phase terminates at an iterate
that has a sufficiently small value of h and is compatible with the filter.
We now present a framework for filter methods that assumes that iterates are generated
by a trust-region method; see Section 18.5 for a discussion of trust-region methods for
constrained optimization.
Algorithm 15.1 (General Filter Method).
Choose a starting point x0 and an initial trust-region radius 0 ;
Set k ← 0;
repeat until a convergence test is satisfied
if the step-generation subproblem is infeasible
Compute xk+1 using the feasibility restoration phase;
else
Compute a trial iterate x + xk + pk ;
if ( f + , h + ) is acceptable to the filter
Set xk+1 x + and add ( f k+1 , h k+1 ) to the filter;
Choose k+1 such that k+1 ≥ k ;
Remove all pairs from the filter that are dominated
by ( f k+1 , h k+1 );
else
Reject the step, set xk+1 xk ;
Choose k+1 < k ;
end if
end if
k ← k + 1;
end repeat
Other enhancements of this simple filter framework are used in practice; they depend
on the choice of algorithm and will be discussed in subsequent chapters.
15.5
THE MARATOS EFFECT
Some algorithms based on merit functions or filters may fail to converge rapidly because they
reject steps that make good progress toward a solution. This undesirable phenomenon is
often called the Maratos effect, because it was first observed by Maratos [199]. It is illustrated
by the following example, in which steps pk , which would yield quadratic convergence if
accepted, cause an increase both in the objective function value and the constraint violation.
15.5.
THE MARATOS EFFECT
contours of f
xk
θ
x*
xk + pk
Figure 15.8
Maratos Effect: Example 15.4.
Note that the constraint is no
longer satisfied after the step
from xk to xk + pk , and the
objective value has increased.
constraint x12 + x 22 = 1
❏ EXAMPLE 15.4
(POWELL [255])
Consider the problem
min f (x1 , x2 ) 2(x12 + x22 − 1) − x1 ,
subject to
x12 + x22 − 1 0.
(15.34)
One can verify (see Figure 15.8) that the optimal solution is x ∗ (1, 0)T , that the
corresponding Lagrange multiplier is λ∗ 32 , and that ∇x2x L(x ∗ , λ∗ ) I .
Let us consider an iterate xk of the form xk (cos θ, sin θ )T , which is feasible for any
value of θ . Suppose that our algorithm computes the following step:
pk sin2 θ
− sin θ cos θ
,
(15.35)
which yields a trial point
x k + pk cos θ + sin2 θ
sin θ (1 − cos θ )
.
By using elementary trigonometric identities, we have that
xk + pk − x ∗ 2 2 sin2 (θ/2),
xk − x ∗ 2 2| sin(θ/2)|,
441
442
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
and therefore
xk + pk − x ∗ 2
1
.
2
∗
2
xk − x 2
Hence, this step approaches the solution at a rate consistent with Q-quadratic convergence.
However, we have that
f (xk + pk ) sin2 θ − cos θ > − cos θ f (xk ),
c(xk + pk ) sin2 θ > c(xk ) 0,
so that, as can be seen in Figure 15.8, both the objective function value and the constraint
violation increase over this step. This behavior occurs for any nonzero value of θ , even if the
initial point is arbitrarily close to the solution.
❐
On the example above, any algorithm that requires reduction of a merit function of
the form
φ(x; µ) f (x) + µh(c(x)),
where h(·) is a nonnegative function satisfying h(0) 0, will reject the good step (15.35).
(Examples of such merit functions include the φ1 and φ2 penalty functions.) The step
(15.35) will also be rejected by the filter mechanism described above because the pair
( f (xk + pk ), h(xk + pk )) is dominated by ( f k , h k ). Therefore, all these approaches will
suffer from the Maratos effect.
If no remedial measures are taken, the Maratos effect can slow optimization methods by interfering with good steps away from the solution and by preventing superlinear
convergence. Strategies for avoiding the Maratos effect include the following.
1. We can use a merit function that does not suffer from the Maratos effect. An example is
Fletcher’s augmented Lagrangian function (15.26).
2. We can use a second-order correction in which we add to pk a step p̂k , which is computed
at c(xk + pk ) and which decreases the constraint violation.
3. We can allow the merit function φ to increase on certain iterations; that is, we can use a
nonmonotone strategy.
We discuss the last two approaches in the next section.
15.6.
15.6
SECOND-ORDER CORRECTION
AND
NONMONOTONE TECHNIQUES
SECOND-ORDER CORRECTION AND NONMONOTONE
TECHNIQUES
By adding a correction term that decreases the constraint violation, various algorithms
are able to overcome the difficulties associated with the Maratos effect. We describe this
technique with respect to the equality-constrained problem, in which the constraint is
c(x) 0, where c : IRn → IR|E| .
Given a step pk , the second-order correction step p̂k is defined to be
p̂k −AkT (Ak AkT )−1 c(xk + pk ),
(15.36)
where Ak A(xk ) is the Jacobian of c at xk . Note that p̂k has the property that it satisfies a
linearization of the constraint c at the point xk + pk , that is,
Ak p̂k + c(xk + pk ) 0.
In fact, p̂k is the minimum-norm solution of this equation. (A different interpretation of
the second-order correction is given in Section 18.3.)
The effect of the correction step p̂k is to decrease the quantity c(x) to the order
of xk − x ∗ 3 , provided the primary step pk satisfies Ak pk + c(xk ) 0. This estimate
indicates that the step from from xk to xk + pk + p̂k will decrease the merit function, at
least near the solution. The cost of this enhancement includes the additional evaluation of
the constraint function c at xk + pk and the linear algebra required to calculate the step p̂k
from (15.36).
We now describe an algorithm that uses a merit function together with a line-search
strategy and a second-order correction step. We assume that the search direction pk and the
penalty parameter µk are computed so that pk is a descent direction for the merit function,
that is, D(φ(xk ; µ); pk ) < 0. In Chapters 18 and 19, we discuss how to accomplish these
goals. The key feature of the algorithm is that, if the full step αk 1 does not produce
satisfactory descent in the merit function, we try the second-order correction step before
backtracking along the original direction pk .
Algorithm 15.2 (Generic Algorithm with Second-Order Correction).
Choose parameters η ∈ (0, 0.5) and τ1 , τ2 with 0 < τ1 < τ2 < 1;
Choose initial point x0 ; set k ← 0;
repeat until a convergence test is satisfied:
Compute a search direction pk ;
Set αk ← 1, newpoint ← false ;
while newpoint = false
if φ(xk + αk pk ; µ) ≤ φ(xk ; µ) + ηαk D(φ(xk ; µ); pk )
Set xk+1 ← xk + αk pk ;
Set newpoint ← true;
443
444
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
else if αk 1
Compute p̂k from (15.36);
if φ(xk + pk + p̂k ; µ) ≤ φ(xk ; µ) + ηD(φ(xk ; µ); pk )
Set xk+1 ← xk + pk + p̂k ;
Set newpoint ← true;
else
Choose new αk in [τ1 αk , τ2 αk ];
end
else
Choose new αk in [τ1 αk , τ2 αk ];
end
end while
end repeat
In this algorithm, the full second-order correction step p̂k is discarded if does not
produce a reduction in the merit function. We do not backtrack along the direction
pk + p̂k because it is not guaranteed to be a descent direction for the merit function. A variation of this algorithm applies the second-order correction step only if the
sufficient decrease condition (15.29) is violated as a result of an increase in the norm of the
constraints.
The second-order correction strategy is effective in practice. The cost of performing
the extra constraint function evaluation and an additional backsolve in (15.36) is outweighed
by added robustness and efficiency.
NONMONOTONE (WATCHDOG) STRATEGY
The inefficiencies caused by the Maratos effect can also be avoided by occasionally
accepting steps that increase the merit function; such steps are called relaxed steps. There
is a limit to our tolerance, however. If a sufficient reduction of the merit function has not
been obtained within a certain number of iterates of the relaxed step (tˆ iterates, say), then
we return to the iterate before the relaxed step and perform a normal iteration, using a line
search or some other technique to force a reduction in the merit function.
In contrast with the second-order correction, which aims only to improve satisfaction
of the constraints, this nonmonotone strategy always takes regular steps pk of the algorithm
that aim both for improved feasibility and optimality. The hope is that any increase in the
merit function over a single step will be temporary, and that subsequent steps will more
than compensate for it.
We now describe a particular instance of the nonmonotone approach called the
watchdog strategy. We set tˆ 1, so that we allow the merit function to increase on just
a single step before insisting on a sufficient decrease in the merit function. As above, we
focus our discussion on a line search algorithm that uses a nonsmooth merit function φ.
We assume that the penalty parameter µ is not changed until a successful cycle has been
15.6.
SECOND-ORDER CORRECTION
AND
NONMONOTONE TECHNIQUES
completed. To simplify the notation, we omit the dependence of φ on µ and write the merit
function as φ(x) and the directional derivative as D(φ(x); pk ).
Algorithm 15.3 (Watchdog).
Choose a constant η ∈ (0, 0.5) and an initial point x0 ;
Set k ← 0, S ← {0};
repeat until a termination test is satisfied
Compute a step pk ;
Set xk+1 ← xk + pk ;
if φ(xk+1 ) ≤ φ(xk ) + ηD(φ(xk ); pk )
k ← k + 1, S ← S ∪ {k};
else
Compute a search direction pk+1 from xk+1 ;
Find αk+1 such that
φ(xk+2 ) ≤ φ(xk+1 ) + ηαk+1 D(φ(xk+1 ); pk+1 );
Set xk+2 ← xk+1 + αk+1 pk+1 ;
if φ(xk+1 ) ≤ φ(xk ) or φ(xk+2 ) ≤ φ(xk ) + ηD(φ(xk ); pk )
k ← k + 2, S ← S ∪ {k};
else if φ(xk+2 ) > φ(xk )
(* return to xk and search along pk *)
Find αk such that φ(xk+3 ) ≤ φ(xk ) + ηαk D(φ(xk ); pk );
Compute xk+3 xk + αk pk ;
k ← k + 3, S ← S ∪ {k};
else
Compute a direction pk+2 from xk+2 ;
Find αk+2 such that
φ(xk+3 ) ≤ φ(xk+2 ) + ηαk+2 D(φ(xk+2 ); pk+2 );
Set xk+3 ← xk+2 + αk+2 pk+2 ;
k ← k + 3, S ← S ∪ {k};
end
end
end (repeat)
The set S is not required by the algorithm and is introduced only to identify the iterates
for which a sufficient merit function reduction was obtained. Note that at least a third of
the iterates have their indices in S. By using this fact, one can show that various constrained
optimization methods that use the watchdog technique are globally convergent. One can
also show that for all sufficiently large k, the step length is αk 1 and the convergence rate
is superlinear.
In practice, it may be advantageous to allow increases in the merit function for more
than one iteration. Values of tˆ such as 5 or 8 are typical. As this discussion indicates, careful implementations of the watchdog technique have a certain degree of complexity, but
445
446
CHAPTER 15.
FUNDAMENTALS
OF
CONSTRAINED ALGORITHMS
the added complexity is worthwhile because the approach has good practical performance.
A potential advantage of the watchdog technique over the second-order correction strategy is that it may require fewer evaluations of the constraint functions. In the best case,
most of the steps will be full steps, and there will rarely be a need to return to an earlier
point.
NOTES AND REFERENCES
Techniques for eliminating linear constraints are described, for example, in Fletcher
[101] and Gill, Murray, and Wright [131]. For a thorough discussion of merit functions see
Boggs and Tolle [33] and Conn, Gould, and Toint [74]. Some of the earliest references on
nonmonotone methods include Grippo, Lampariello and Lucidi [158], and Chamberlain et
al [57]; see [74] for a review of nonmonotone techniques and an extensive list of references.
The concept of a filter was introduced by Fletcher and Leyffer [105]; our discussion of
filters is based on that paper. Second-order correction steps are motivated and discussed in
Fletcher [101].
✐
EXERCISES
✐
15.1 In Example 15.1, consider these three choices of the working set: W {3},
W {1, 2}, W {2, 3}. Show that none of these working sets are the optimal active set for
(15.5).
✐
15.2 For the problem in Example 15.3, perform simple elimination of the variables
x2 and x5 to obtain an unconstrained problem in the remaining variables x1 , x3 , x4 , and
x6 . Similarly to (15.12), express the eliminated variables explicitly in terms of the retained
variables.
✐
15.3 Do the following problems have solutions? Explain.
min x1 + x2
subject to x12 + x22 2, 0 ≤ x1 ≤ 1, 0 ≤ x2 ≤ 1;
min x1 + x2
subject to x12 + x22 ≤ 1, x1 + x2 3;
min x1 x2
subject to x1 + x2 2.
✐
15.4 Show that if in Example 15.2 we eliminate x in terms of y, then the correct
solution of the problem is obtained by performing unconstrained minimization.
✐
✐
15.5 Show that the basis matrices (15.15) are linearly independent.
15.6 Show that the particular solution Q 1 R −T T b of Ax b is identical to
A (A A T )−1 b.
T
15.6.
SECOND-ORDER CORRECTION
AND
NONMONOTONE TECHNIQUES
✐
15.7 In this exercise we compute basis matrices that attempt to compromise between
the orthonormal basis (15.22) and simple elimination (15.15). We assume that the basis
matrix is given by the first m columns of A, so that P I in (15.7), and define
Y I
(B −1 N )T
,
Z
−B −1 N
I
.
(15.37)
(a) Show that the columns of Y and Z are no longer of norm 1 and that the relations
AZ 0 and Y T Z 0 hold. Therefore, the columns of Y and Z form a linearly
independent set, showing that (15.37) is a valid choice of the basis matrices.
(b) Show that the particular solution Y (AY )−1 b defined by this choice of Y is, as in the
orthogonal factorization approach, the minimum-norm solution of Ax b. More
specifically, show that
Y (AY )−1 A T (A A T )−1 .
It follows that the matrix Y (AY )−1 is independent of the choice of basis matrix B in
(15.7), and its conditioning is determined by that of A alone. (Note, however, that the
matrix Z still depends explicitly on B, so a careful choice of B is needed to ensure well
conditioning in this part of the computation.)
✐
15.8 Verify that by adding the inequality constraint 3x1 + 2x3 ≥ 1 to the problem
(15.11), the elimination (15.12) transforms the problem into one of minimizing the function
(15.13) subject to the inequality constraint (15.23).
447
This is pag
Printer: O
16
CHAPTER
Quadratic
Programming
An optimization problem with a quadratic objective function and linear constraints is called
a quadratic program. Problems of this type are important in their own right, and they also
arise as subproblems in methods for general constrained optimization, such as sequential
quadratic programming (Chapter 18), augmented Lagrangian methods (Chapter 17), and
interior-point methods (Chapter 19).
CHAPTER 16.
QUADRATIC PROGRAMMING
The general quadratic program (QP) can be stated as
min
x
subject to
q(x) 12 x T Gx + x T c
(16.1a)
aiT x bi ,
i ∈ E,
(16.1b)
≥ bi ,
i ∈ I,
(16.1c)
aiT x
where G is a symmetric n × n matrix, E and I are finite sets of indices, and c, x, and
{ai }, i ∈ E ∪ I, are vectors in IRn . Quadratic programs can always be solved (or shown
to be infeasible) in a finite amount of computation, but the effort required to find a
solution depends strongly on the characteristics of the objective function and the number
of inequality constraints. If the Hessian matrix G is positive semidefinite, we say that (16.1)
is a convex QP, and in this case the problem is often similar in difficulty to a linear program.
(Strictly convex QPs are those in which G is positive definite.) Nonconvex QPs, in which G
is an indefinite matrix, can be more challenging because they can have several stationary
points and local minima.
In this chapter we focus primarily on convex quadratic programs. We start by
considering an interesting application of quadratic programming.
❏ EXAMPLE 16.1
(PORTFOLIO OPTIMIZATION)
Every investor knows that there is a tradeoff between risk and return: To increase the
expected return on investment, an investor must be willing to tolerate greater risks. Portfolio
theory studies how to model this tradeoff given a collection of n possible investments with
returns ri , i 1, 2, . . . , n. The returns ri are usually not known in advance and are often
assumed to be random variables that follow a normal distribution. We can characterize these
variables by their expected value µi E[ri ] and their variance σi2 E[(ri − µi )2 ]. The
variance measures the fluctuations of the variable ri about its mean, so that larger values
of σi indicate riskier investments. The returns are not in general independent, and we can
define correlations between pairs of returns as follows:
ρi j E[(ri − µi )(r j − µ j )]
,
σi σ j
for i, j 1, 2, . . . , n.
The correlation measures the tendency of the return on investments i and j to move in the
same direction. Two investments whose returns tend to rise and fall together have a positive
correlation; the nearer ρi j is to 1, the more closely the two investments track each other.
Investments whose returns tend to move in opposite directions have a negative correlation.
An investor constructs a portfolio by putting a fraction xi of the available funds into
investment i, for i 1, 2, . . . , n. Assuming that all available funds are invested and that
n
xi 1 and x ≥ 0. The return on the
short-selling is not allowed, the constraints are i1
449
450
CHAPTER 16.
QUADRATIC PROGRAMMING
portfolio is given by
R
n
x i ri .
(16.2)
i1
To measure the desirability of the portfolio, we need to obtain measures of its expected
return and variance. The expected return is simply
E[R] E
n
x i ri
i1
n
xi E[ri ] x T µ,
i1
while the variance is given by
Var[R] E[(R − E[R]) ] 2
n
n xi x j σi σ j ρi j x T Gx,
i1 j1
where the n × n symmetric positive semidefinite matrix G defined by
G i j ρi j σi σ j
is called the covariance matrix.
Ideally, we would like to find a portfolio for which the expected return x T µ is large
while the variance x T Gx is small. In the model proposed by Markowitz [201], we combine
these two aims into a single objective function with the aid of a “risk tolerance parameter”
denoted by κ, and we solve the following problem to find the optimal portfolio:
max x T µ − κ x T Gx,
subject to
n
xi 1, x ≥ 0.
i1
The value chosen for the nonnegative parameter κ depends on the preferences of the
individual investor. Conservative investors, who place more emphasis on minimizing risk
in their portfolio, would choose a large value of κ to increase the weight of the variance
measure in the objective function. More daring investors, who are prepared to take on more
risk in the hope of a higher expected return, would choose a smaller value of κ.
The difficulty in applying this portfolio optimization technique to real-life investing
lies in defining the expected returns, variances, and correlations for the investments in
question. Financial professionals often combine historical data with their own insights and
expectations to produce values of these quantities.
❐
16.1.
16.1
EQUALITY-CONSTRAINED QUADRATIC PROGRAMS
EQUALITY-CONSTRAINED QUADRATIC PROGRAMS
We begin our discussion of algorithms for quadratic programming by considering the
case in which only equality constraints are present. Techniques for this special case are
applicable also to problems with inequality constraints since, as we see later in this chapter,
some algorithms for general QP require the solution of an equality-constrained QP at each
iteration.
PROPERTIES OF EQUALITY-CONSTRAINED QPs
For simplicity, we write the equality constraints in matrix form and state the equalityconstrained QP as follows:
min
x
subject to
def
q(x) 12 x T Gx + x T c
(16.3a)
Ax b,
(16.3b)
where A is the m × n Jacobian of constraints (with m ≤ n) whose rows are aiT , i ∈ E and
b is the vector in IRm whose components are bi , i ∈ E. For the present, we assume that A
has full row rank (rank m) so that the constraints (16.3b) are consistent. (In Section 16.8
we discuss the case in which A is rank deficient.)
The first-order necessary conditions for x ∗ to be a solution of (16.3) state that there is
a vector λ∗ such that the following system of equations is satisfied:
G
−A T
A
0
x∗
λ∗
−c
.
b
(16.4)
These conditions are a consequence of the general result for first-order optimality conditions,
Theorem 12.1. As in Chapter 12, we call λ∗ the vector of Lagrange multipliers. The system
(16.4) can be rewritten in a form that is useful for computation by expressing x ∗ as x ∗ x + p, where x is some estimate of the solution and p is the desired step. By introducing
this notation and rearranging the equations, we obtain
G
AT
A
0
−p
λ∗
g
h
,
(16.5)
where
h Ax − b,
g c + Gx,
p x ∗ − x.
(16.6)
The matrix in (16.5) is called the Karush–Kuhn–Tucker (KKT) matrix, and the following result gives conditions under which it is nonsingular. As in Chapter 15, we use Z to
451
452
CHAPTER 16.
QUADRATIC PROGRAMMING
denote the n × (n − m) matrix whose columns are a basis for the null space of A. That is,
Z has full rank and satisfies AZ 0.
Lemma 16.1.
Let A have full row rank, and assume that the reduced-Hessian matrix Z T G Z is positive
definite. Then the KKT matrix
K G
AT
A
0
(16.7)
is nonsingular, and hence there is a unique vector pair (x ∗ , λ∗ ) satisfying (16.4).
PROOF. Suppose there are vectors w and v such that
G
AT
A
0
w
0.
v
(16.8)
Since Aw 0, we have from (16.8) that
0
w
T v
G
AT
A
0
w
v
w T Gw.
Since w lies in the null space of A, it can be written as w Z u for some vector u ∈ IRn−m .
Therefore, we have
0 w T Gw u T Z T G Z u,
which by positive definiteness of Z T G Z implies that u 0. Therefore, w 0, and by
(16.8), A T v 0. Full row rank of A then implies that v 0. We conclude that equation
(16.8) is satisfied only if w 0 and v 0, so the matrix is nonsingular, as claimed.
❏ EXAMPLE 16.2
Consider the quadratic programming problem
min q(x) 3x12 + 2x1 x2 + x1 x3 + 2.5x22 + 2x2 x3 + 2x32 − 8x1 − 3x2 − 3x3 ,
subject to
x1 + x3 3,
x2 + x3 0.
(16.9)
16.1.
EQUALITY-CONSTRAINED QUADRATIC PROGRAMS
We can write this problem in the form (16.3) by defining
⎡
6
⎢
G⎣ 2
1
2
1
⎡
⎤
5
⎥
2 ⎦,
2
4
−8
⎤
⎥
⎢
c ⎣ −3 ⎦ ,
A
−3
1
0
1
0
1
1
,
b
3
0
.
The solution x ∗ and optimal Lagrange multiplier vector λ∗ are given by
x ∗ (2, −1, 1)T ,
λ∗ (3, −2)T .
In this example, the matrix G is positive definite, and the null-space basis matrix can be
defined as in (15.15), giving
Z (−1, −1, 1)T .
(16.10)
❐
We have seen that when the conditions of Lemma 16.1 are satisfied, there is a unique
vector pair (x ∗ , λ∗ ) that satisfies the first-order necessary conditions for (16.3). In fact, the
second-order sufficient conditions (see Theorem 12.6) are also satisfied at (x ∗ , λ∗ ), so x ∗ is
a strict local minimizer of (16.3). In fact, we can use a direct argument to show that x ∗ is a
global solution of (16.3).
Theorem 16.2.
Let A have full row rank and assume that the reduced-Hessian matrix Z T G Z is positive
definite. Then the vector x ∗ satisfying (16.4) is the unique global solution of (16.3).
PROOF. Let x be any other feasible point (satisfying Ax b), and as before, let p denote
the difference x ∗ − x. Since Ax ∗ Ax b, we have that Ap 0. By substituting into the
objective function (16.3a), we obtain
q(x) 12 (x ∗ − p)T G(x ∗ − p) + c T (x ∗ − p)
12 p T Gp − p T Gx ∗ − c T p + q(x ∗ ).
From (16.4) we have that Gx ∗ −c + A T λ∗ , so from Ap 0 we have that
p T Gx ∗ p T (−c + A T λ∗ ) − p T c.
By substituting this relation into (16.11), we obtain
q(x) 12 p T Gp + q(x ∗ ).
(16.11)
453
454
CHAPTER 16.
QUADRATIC PROGRAMMING
Since p lies in the null space of A, we can write p Z u for some vector u ∈ IRn−m , so that
q(x) 12 u T Z T G Z u + q(x ∗ ).
By positive definiteness of Z T G Z , we conclude that q(x) > q(x ∗ ) except when u 0, that
is, when x x ∗ . Therefore, x ∗ is the unique global solution of (16.3).
When the reduced Hessian matrix Z T G Z is positive semidefinite with zero eigenvalues, the vector x ∗ satisfying (16.4) is a local minimizer but not a strict local minimizer. If
the reduced Hessian has negative eigenvalues, then x ∗ is only a stationary point, not a local
minimizer.
16.2
DIRECT SOLUTION OF THE KKT SYSTEM
In this section we discuss efficient methods for solving the KKT system (16.5). The first
important observation is that if m ≥ 1, the KKT matrix is always indefinite. We define the
inertia of a symmetric matrix K to be the scalar triple that indicates the numbers n + , n − ,
and n 0 of positive, negative, and zero eigenvalues, respectively, that is,
inertia(K ) (n + , n − , n 0 ).
The following result characterizes the inertia of the KKT matrix.
Theorem 16.3.
Let K be defined by (16.7), and suppose that A has rank m. Then
inertia(K ) inertia(Z T G Z ) + (m, m, 0).
Therefore, if Z T G Z is positive definite, inertia(K ) (n, m, 0).
The proof of this result is given in [111], for example. Note that the assumptions of
this theorem are satisfied by Example 16.2. Hence, if we construct the 5 × 5 matrix K using
the data of this example, we obtain inertia(K ) (3, 2, 0).
Knowing that the KKT system is indefinite, we now describe the main direct techniques
used to solve (16.5).
FACTORING THE FULL KKT SYSTEM
One option for solving (16.5) is to perform a triangular factorization on the full KKT
matrix and then perform backward and forward substitution with the triangular factors.
Because of indefiniteness, we cannot use the Cholesky factorization. We could use Gaussian
16.2.
DIRECT SOLUTION
OF THE
KKT SYSTEM
elimination with partial pivoting (or a sparse variant thereof) to obtain the L and U factors,
but this approach has the disadvantage that it ignores the symmetry.
The most effective strategy in this case is to use a symmetric indefinite factorization,
which we have discussed in Chapter 3 and the Appendix. For a general symmetric matrix
K , this factorization has the form
PT K P L BLT ,
(16.12)
where P is a permutation matrix, L is unit lower triangular, and B is block-diagonal with
either 1 × 1 or 2 × 2 blocks. The symmetric permutations defined by the matrix P are
introduced for numerical stability of the computation and, in the case of large sparse K ,
for maintaining sparsity. The computational cost of the symmetric indefinite factorization
(16.12) is typically about half the cost of sparse Gaussian elimination.
To solve (16.5), we first compute the factorization (16.12) of the coefficient matrix.
We then perform the following sequence of operations to arrive at the solution:
solve Lz P
T
solve B ẑ z
solve L z̄ ẑ
−p
set
P z̄.
λ∗
T
g
h
to obtain z;
to obtain ŷ;
to obtain z̄;
Since multiplications with the permutation matrices P and P T can be performed by simply
rearranging vector components, they are inexpensive. Solution of the system B ẑ z entails
solving a number of small 1 × 1 and 2 × 2 systems, so the number of operations is a small
multiple of the system dimension (m + n), again inexpensive. Triangular substitutions with
L and L T are more costly. Their precise cost depends on the amount of sparsity, but is
usually significantly less than the cost of performing the factorization (16.12).
This approach of factoring the full (n + m) × (n + m) KKT matrix (16.7) is quite
effective on many problems. It may be expensive, however, when the heuristics for choosing
the permutation matrix P are not able to maintain sparsity in the L factor, so that L becomes
much more dense than the original coefficient matrix.
SCHUR-COMPLEMENT METHOD
Assuming that G is positive definite, we can multiply the first equation in (16.5) by
AG −1 and then subtract the second equation to obtain a linear system in the vector λ∗ alone:
(AG −1 A T )λ∗ (AG −1 g − h).
(16.13)
455
456
CHAPTER 16.
QUADRATIC PROGRAMMING
We solve this symmetric positive definite system for λ∗ and then recover p from the first
equation in (16.5) by solving
Gp A T λ∗ − g.
(16.14)
This approach requires us to perform operations with G −1 , as well as to compute the
factorization of the m × m matrix AG −1 A T . Therefore, it is most useful when:
• G is well conditioned and easy to invert (for instance, when G is diagonal or blockdiagonal); or
• G −1 is known explicitly through a quasi-Newton updating formula; or
• the number of equality constraints m is small, so that the number of backsolves needed
to form the matrix AG −1 A T is not too large.
The name “Schur-Complement method” derives from the fact that, by applying block
Gaussian elimination to (16.7) using G as the pivot, we obtain the block upper triangular
system
G
AT
0
−AG −1 A T
.
(16.15)
In linear algebra terminology, the matrix AG −1 A T is the Schur complement of G in the
matrix K of (16.7). By applying this block elimination technique to the system (16.5), and
performing a block backsolve, we obtain (16.13), (16.14).
We can use an approach like the Schur-complement method to derive an explicit
inverse formula for the KKT matrix in (16.5). This formula is
G
AT
A
0
−1
C
E
ET
F
,
(16.16)
with
C G −1 − G −1 A T (AG −1 A T )−1 AG −1 ,
E G −1 A T (AG −1 A T )−1 ,
F −(AG −1 A T )−1 .
The solution of (16.5) can be obtained by multiplying its right-hand side by this inverse
matrix. If we take advantage of common expressions, and group the terms appropriately,
we recover the approach (16.13), (16.14).
16.2.
DIRECT SOLUTION
OF THE
KKT SYSTEM
NULL-SPACE METHOD
The null-space method does not require nonsingularity of G and therefore has wider
applicability than the Schur-complement method. It assumes only that the conditions of
Lemma 16.1 hold, namely, that A has full row rank and that Z T G Z is positive definite.
However, it requires knowledge of the null-space basis matrix Z . Like the Schur-complement
method, it exploits the block structure in the KKT system to decouple (16.5) into two smaller
systems.
Suppose that we partition the vector p in (16.5) into two components, as follows:
p Y pY + Z pZ ,
(16.17)
where Z is the n × (n − m) null-space matrix, Y is any n × m matrix such that [Y | Z ]
is nonsingular, pY is an m-vector, and pZ is an (n − m)-vector. The matrices Y and Z
were discussed in Section 15.3, where Figure 15.4 shows that Y xY is a particular solution of
Ax b, while Z xZ is a displacement along these constraints.
By substituting p into the second equation of (16.5) and recalling that AZ 0, we
obtain
(AY ) pY −h.
(16.18)
Since A has rank m and [Y | Z ] is n × n nonsingular, the product A[Y | Z ] [AY | 0] has
rank m. Therefore, AY is a nonsingular m × m matrix, and pY is well determined by the
equations (16.18). Meanwhile, we can substitute (16.17) into the first equation of (16.5) to
obtain
−GY pY − G Z pZ + A T λ∗ g
and multiply by Z T to obtain
(Z T G Z ) pZ −Z T GY pY − Z T g.
(16.19)
This system can be solved by performing a Cholesky factorization of the reduced-Hessian
matrix Z T G Z to determine pZ . We therefore can compute the total step p Y pY + Z pZ .
To obtain the Lagrange multiplier, we multiply the first block row in (16.5) by Y T to obtain
the linear system
(AY )T λ∗ Y T (g + Gp),
which can be solved for λ∗ .
(16.20)
457
458
CHAPTER 16.
QUADRATIC PROGRAMMING
❏ EXAMPLE 16.3
Consider the problem (16.9) given in Example 16.2. We can choose
⎡
2/3
−1/3
1/3
1/3
⎢
Y ⎣ −1/3
⎤
⎥
2/3 ⎦
and set Z as in (16.10). Note that AY I .
Suppose we have x (0, 0, 0)T in (16.6). Then
⎡
−8
⎤
⎢
⎥
g c + Gx c ⎣ −3 ⎦ .
h Ax − b −b,
−3
Simple calculation shows that
3
pY ,
0
pZ ,
0
so that
⎡
2
⎤
⎢
⎥
p x ∗ − x Y pY + Z pZ ⎣ −1 ⎦ .
1
After recovering λ∗ from (16.20), we conclude that
⎡
2
⎤
⎢
⎥
x ∗ ⎣ −1 ⎦ ,
1
λ∗ 3
−2
.
❐
The null-space approach can be very effective when the number of degrees of freedom
n − m is small. Its main limitation lies in the need for the null-space matrix Z which, as we
have seen in Chapter 15, can be expensive to compute in some large problems. The matrix Z
is not uniquely defined and, if it is poorly chosen, the reduced system (16.19) may become ill
conditioned. If we choose Z to have orthonormal columns, as is normally done in software
for small and medium-sized problems, then the conditioning of Z T G Z is at least as good
as that of G itself. When A is large and sparse, however, an orthonormal Z is expensive to
16.3.
ITERATIVE SOLUTION
OF THE
KKT SYSTEM
compute, so for practical reasons we are often forced to use one of the less reliable choices
of Z described in Chapter 15.
It is difficult to give hard and fast rules about the relative effectiveness of null-space and
Schur-complement methods, because factors such as fill-in during computation of Z vary
significantly even among problems of the same dimension. In general, we can recommend
the Schur-complement method if G is positive definite and AG −1 A T can be computed
relatively cheaply (because G is easy to invert or because m is small relative to n). Otherwise,
the null-space method is often preferable, in particular when it is much more expensive to
compute factors of G than to compute the null-space matrix Z and the factors of Z T G Z .
16.3
ITERATIVE SOLUTION OF THE KKT SYSTEM
An alternative to the direct factorization techniques discussed in the previous section is to
use an iterative method to solve the KKT system (16.5). Iterative methods are suitable for
solving very large systems and often lend themselves well to parallelization. The conjugate
gradient (CG) method is not recommended for solving the full system (16.5) as written,
because it can be unstable on systems that are not positive definite. Better options are
Krylov methods for general linear or symmetric indefinite systems. Candidates include
the GMRES, QMR, and LSQR methods; see the Notes and References at the end of the
chapter. Other iterative methods can be derived from the null-space approach by applying
the conjugate gradient method to the reduced system (16.19). Methods of this type are key
to the algorithms of Chapters 18 and 19, and are discussed in the remainder of this section.
We assume throughout that Z T G Z is positive definite.
CG APPLIED TO THE REDUCED SYSTEM
We begin our discussion of iterative null-space methods by deriving the underlying
equations in the notation of the equality-constrained QP (16.3). Expressing the solution of
the quadratic program (16.3) as
x ∗ Y xY + Z xZ ,
(16.21)
for some vectors xZ ∈ IRn−m , xY ∈ IRm , the constraints Ax b yield
AY xY b,
(16.22)
which determines the vector xY . In Chapter 15, various practical choices of Y are described,
some of which allow (16.22) to be solved economically. Substituting (16.21) into (16.3), we
see that xZ solves the unconstrained reduced problem
min 12 xZ T Z T G Z xZ + xZ T cZ ,
xZ
459
460
CHAPTER 16.
QUADRATIC PROGRAMMING
where
cZ Z T GY xY + Z T c.
(16.23)
The solution xZ satisfies the linear system
Z T G Z xZ −cZ .
(16.24)
Since Z T G Z is positive definite, we can apply the CG method to this linear system and
substitute xZ into (16.21) to obtain a solution of (16.3).
As discussed in Chapter 5, preconditioning can improve the rate of convergence of
the CG iteration, so we assume that a preconditioner WZZ is given. The preconditioned CG
method (Algorithm 5.3) applied to the (n − m)-dimensional reduced system (16.24) is as
follows. (We denote the steps produced by the CG iteration by dZ .)
Algorithm 16.1 (Preconditioned CG for Reduced Systems).
Choose an initial point xZ ;
Compute rZ Z T G Z xZ + cZ , gZ WZZ −1rZ , and dZ −gZ ;
repeat
α ← rZ T gZ /dZ T Z T G Z dZ ;
(16.25a)
xZ ← xZ + αdZ ;
(16.25b)
T
rZ ← rZ + α Z G Z dZ ;
(16.25c)
gZ ← WZZ rZ ;
(16.25d)
+
−1
+
+
β ← (rZ + )T gZ + /rZ T gZ ;
(16.25e)
dZ ← −gZ + βdZ ;
(16.25f)
+
gZ ← gZ ;
+
rZ ← rZ ;
+
(16.25g)
until a termination test is satisfied.
This iteration may be terminated when, for example, rZ T WZZ −1rZ is sufficiently small.
In this approach, it is not necessary to form the reduced Hessian Z T G Z explicitly
because the CG method requires only that we compute matrix-vector products involving
this matrix. In fact, it is not even necessary to form Z explicitly as long as we are able to
compute products of Z and Z T with arbitrary vectors. For some choices of Z , these products
are much cheaper to compute than Z itself, as we have seen in Chapter 15.
The preconditioner WZZ is a symmetric, positive definite matrix of dimension n − m,
which might be chosen to cluster the eigenvalues of WZZ −1/2 (Z T G Z )WZZ −1/2 and to reduce
the span between the smallest and largest eigenvalues. An ideal choice of preconditioner is
one for which WZZ −1/2 (Z T G Z )WZZ −1/2 I , that is, WZZ Z T G Z . Motivated by this ideal,
we consider preconditioners of the form
WZZ Z T H Z ,
(16.26)
16.3.
ITERATIVE SOLUTION
OF THE
KKT SYSTEM
where H is a symmetric matrix such that Z T H Z is positive definite. Some choices of H are
discussed below. Preconditioners of the form (16.26) allow us to apply the CG method in
n-dimensional space, as we discuss next.
THE PROJECTED CG METHOD
It is possible to design a modification of the Algorithm 16.1 that avoids operating with
the null-space basis Z , provided we use a preconditioner of the form (16.26) and a particular
solution of the equation Ax b. This approach works implicitly with an orthogonal matrix
Z and is not affected by ill conditioning in A or by a poor choice of Z .
After the solution xZ of (16.24) has been computed by using Algorithm 16.1, it
must be multiplied by Z and substituted in (16.21) to give the solution of the quadratic
program (16.3). Alternatively, we may rewrite Algorithm 16.1 to work directly with the
vector x Z xZ + Y xY , where the Y xY term is fixed at the start and the xZ term is updated
(implicitly) within each iteration. To specify this form of the CG algorithm, we introduce
the n-vectors x, r , g, and d, which satisfy x Z xZ + Y xY , Z T r rZ , g Z gZ , and d Z dZ ,
respectively. We also define the scaled n × n projection matrix P as follows:
P Z (Z T H Z )−1 Z T ,
(16.27)
where H is the preconditioning matrix from (16.26). The CG iteration in n-dimensional
space can be specified as follows.
Algorithm 16.2 (Projected CG Method).
Choose an initial point x satisfying Ax b;
Compute r Gx + c, g Pr , and d −g;
repeat
α ← r T g/d T Gd;
(16.28a)
x ← x + αd;
(16.28b)
+
r ← r + αGd;
+
(16.28c)
+
g ← Pr ;
+ T
(16.28d)
+
β ← (r ) g /r g;
T
(16.28e)
+
d ← −g + βd;
+
g←g ;
until a convergence test is satisfied.
(16.28f)
+
r ←r ;
(16.28g)
461
462
CHAPTER 16.
QUADRATIC PROGRAMMING
A practical stop test is to terminate when r T g r T Pr is smaller than a prescribed
tolerance.
Note that the vector g + , which we call the preconditioned residual, has been defined to
be in the null space of A. As a result, in exact arithmetic, all the search directions d generated
by Algorithm 16.2 also lie in the null space of A, and thus the iterates x all satisfy Ax b.
It is not difficult to verify (see Exercise 16.14) that the iteration is well defined if Z T G Z
and Z T H Z are positive definite. The reader can also verify that the iterates x generated by
Algorithm 16.2 are related to the iterates xZ of Algorithm 16.1 via (16.21).
Two simple choices of the preconditioning matrix H are H diag(|G ii |) and H I .
In some applications, it is effective to define H as a block diagonal submatrix of G.
Algorithm 16.2 makes use of the null-space basis Z only through the operator (16.27).
It is possible, however, to compute Pr without knowing a representation of the null-space
basis Z . For simplicity, we first consider the case in which H I , so that P is the orthogonal
projection operator onto the null space of A. We use PI to denote this special case of P,
that is,
PI Z (Z T Z )−1 Z T .
(16.29)
The computation of the preconditioned residual g + PIr + in (16.28d) can be performed
in two ways. The first is to express PI by the equivalent formula
PI I − A T (A A T )−1 A
(16.30)
and thus compute g + PIr + . We can then write g + r + − A T v + , where v + is the solution
of the system
A A T v + Ar + .
(16.31)
This approach for computing the projection g + PIr + is called the normal equations approach; the system (16.31) can be solved by using a Cholesky factorization
of A A T .
The second approach is to express the projection (16.28d) as the solution of the
augmented system
I
AT
A
0
g+
v+
r+
0
,
(16.32)
which can be solved by means of a symmetric indefinite factorization, as discussed earlier.
We call this approach the augmented system approach.
16.4.
INEQUALITY-CONSTRAINED PROBLEMS
We suppose now that the preconditioning has the general form of (16.27) and (16.28d).
When H is nonsingular, we can compute g + as follows:
g + Pr + ,
P H −1 I − A T (AH −1 A T )−1 AH −1 .
where
(16.33)
Otherwise, when z T H z 0 for all nonzero z with Az 0, we can find g + as the solution
of the system
H
AT
A
0
g+
v+
r+
.
0
(16.34)
While (16.33) is unappealing when H −1 does not have a simple form, (16.34) is a useful
generalization of (16.32). A “perfect” preconditioner is obtained by taking H G, but
other choices for H are also possible, provided that Z T H Z is positive definite. The matrix
in (16.34) is often called a constraint preconditioner.
None of these procedures for computing the projection makes use of a null-space
basis Z ; only the factorization of matrices involving A is required. Significantly, all these
forms allow us to compute an initial point satisfying Ax b. The operator g + PIr +
relies on a factorization of A A T from which we can compute x A T (A A T )−1 b, while
factorizations of the system matrices in (16.32) and (16.34) allow us to find a suitable x by
solving
I
AT
A
0
x
y
0
b
or
H
AT
A
0
x
y
0
b
.
Therefore we can compute an initial point for Algorithm 16.2 at the cost of one backsolve,
using the factorization of the system needed to perform the projection operators.
We point out that these approaches for computing g + can give rise to significant round-off errors, so the use of iterative refinement is recommended to improve
accuracy.
16.4
INEQUALITY-CONSTRAINED PROBLEMS
In the remainder of the chapter we discuss several classes of algorithms for solving convex
quadratic programs that contain both inequality and equality constraints. Active-set methods
have been widely used since the 1970s and are effective for small- and medium-sized
problems. They allow for efficient detection of unboundedness and infeasibility and typically
return an accurate estimate of the optimal active set. Interior-point methods are more recent,
having become popular in the 1990s. They are well suited for large problems but may not
be the most effective when a series of related QPs must be solved. We also study a special
463
464
CHAPTER 16.
QUADRATIC PROGRAMMING
type of active-set methods called a gradient projection method, which is most effective when
the only constraints in the problem are bounds on the variables.
OPTIMALITY CONDITIONS FOR INEQUALITY-CONSTRAINED PROBLEMS
We begin our discussion with a brief review of the optimality conditions for inequalityconstrained quadratic programming, then discuss some of the less obvious properties of the
solutions.
Theorem 12.1 can be applied to (16.1) by noting that the Lagrangian for this problem
is
L(x, λ) 12 x T Gx + x T c −
λi (aiT x − bi ).
(16.35)
i∈I∪E
As in Definition 12.1, the active set A(x ∗ ) consists of the indices of the constraints for which
equality holds at x ∗ :
3
2
A(x ∗ ) i ∈ E ∪ I | aiT x ∗ bi .
(16.36)
By specializing the KKT conditions (12.34) to this problem, we find that any solution x ∗
of (16.1) satisfies the following first-order conditions, for some Lagrange multipliers λi∗ ,
i ∈ A(x ∗ ):
Gx ∗ + c −
λi∗ ai 0,
(16.37a)
i∈A(x ∗ )
aiT x ∗ bi ,
aiT x ∗
λi∗
≥ bi ,
≥ 0,
for all i ∈ A(x ∗ ),
(16.37b)
∗
for all i ∈ I\A(x ),
(16.37c)
∗
(16.37d)
for all i ∈ I ∩ A(x ).
A technical point: In Theorem 12.1 we assumed that the linear independence constraint qualification (LICQ) was satisfied. As mentioned in Section 12.6, this theorem still
holds if we replace LICQ by other constraint qualifications, such as linearity of the constraints, which is certainly satisfied for quadratic programming. Hence, in the optimality
conditions for quadratic programming given above, we need not assume that the active
constraints are linearly independent at the solution.
For convex QP, when G is positive semidefinite, the conditions (16.37) are in fact
sufficient for x ∗ to be a global solution, as we now prove.
Theorem 16.4.
If x ∗ satisfies the conditions (16.37) for some λi∗ , i ∈ A(x ∗ ), and G is positive semidefinite,
then x ∗ is a global solution of (16.1).
16.4.
INEQUALITY-CONSTRAINED PROBLEMS
PROOF. If x is any other feasible point for (16.1), we have that aiT x bi for all i ∈ E and
aiT x ≥ bi for all i ∈ A(x ∗ ) ∩ I. Hence, aiT (x − x ∗ ) 0 for all i ∈ E and aiT (x − x ∗ ) ≥ 0
for all i ∈ A(x ∗ ) ∩ I. Using these relationships, together with (16.37a) and (16.37d), we
have that
(x − x ∗ )T (Gx ∗ + c) λi∗ aiT (x − x ∗ ) +
i∈E
λi∗ aiT (x − x ∗ ) ≥ 0.
(16.38)
i∈A(x ∗ )∩I
By elementary manipulation, we find that
q(x) q(x ∗ ) + (x − x ∗ )T (Gx ∗ + c) + 12 (x − x ∗ )T G(x − x ∗ )
≥ q(x ∗ ) + 12 (x − x ∗ )T G(x − x ∗ )
≥ q(x ∗ ),
where the first inequality follows from (16.38) and the second inequality follows from
positive semidefiniteness of G. We have shown that q(x) ≥ q(x ∗ ) for any feasible x, so x ∗
is a global solution.
By a trivial modification of this proof, we see that x ∗ is actually the unique global
solution when G is positive definite.
We can also apply the theory from Section 12.5 to derive second-order optimality
conditions for (16.1). Second-order sufficient conditions for x ∗ to be a local minimizer are
satisfied if Z T G Z is positive definite, where Z is defined to be a null-space basis matrix
for the active constraint Jacobian matrix, which is the matrix whose rows are aiT for all
i ∈ A(x ∗ ). In this case, x ∗ is a strict local solution, according to Theorem 12.6.
When G is not positive definite, the general problem (16.1) may have more than one
strict local solution. As mentioned above, such problems are called “nonconvex QPs” or
“indefinite QPs,” and they cause some complications for algorithms. Examples of indefinite
QPs are illustrated in Figure 16.1. On the left we have plotted the feasible region and
the contours of a quadratic objective q(x) in which G has one positive and one negative
eigenvalue. We have indicated by + or − that the function tends toward plus or minus
infinity in that direction. Note that x ∗∗ is a local maximizer, x ∗ a local minimizer, and the
center of the box is a stationary point. The picture on the right in Figure 16.1, in which both
eigenvalues of G are negative, shows a global maximizer at x̃ and local minimizers at x∗ and
x∗∗ .
DEGENERACY
A second property that causes difficulties for some algorithms is degeneracy. Confusingly, this term has been given a variety of meanings. It refers to situations in
which
465
466
CHAPTER 16.
QUADRATIC PROGRAMMING
feasible
region
x
*
+
x**
x
x* -
-
~
x
+
Figure 16.1
x
**
Nonconvex quadratic programs.
x*
x*
Figure 16.2
Degenerate solutions of quadratic programs.
(a) the active constraint gradients ai , i ∈ A(x ∗ ), are linearly dependent at the solution x ∗ ,
and/or
(b) the strict complementarity condition of Definition 12.5 fails to hold, that is, there is
some index i ∈ A(x ∗ ) such that all Lagrange multipliers satisfying (16.37) have λi∗ 0.
(Such constraints are weakly active according to Definition 12.8.)
Two examples of degeneracy are shown in Figure 16.2. In the left-hand picture, there
is a single active constraint at the solution x∗ , which is also an unconstrained minimizer
of the objective function. In the notation of (16.37a), we have that Gx∗ + c 0, so that
16.5.
ACTIVE-SET METHODS
FOR
CONVEX QPS
the lone Lagrange multiplier must be zero. In the right-hand picture, three constraints are
active at the solution x∗ . Since each of the three constraint gradients is a vector in IR2 , they
must be linearly dependent.
Lack of strict complementarity is also illustrated by the problem
min x12 + (x2 + 1)2
subject to x ≥ 0,
which has a solution at x ∗ 0 at which both constraints are active. Strict complementarity
does not hold at x ∗ because the Lagrange multiplier associated with the active constraint
x1 ≥ 0 is zero.
Degeneracy can cause problems for algorithms for two main reasons. First, linear
dependence of the active constraint gradients can cause numerical difficulties in the step
computation because certain matrices that we need to factor become rank deficient. Second,
when the problem contains weakly active constraints, it is difficult for the algorithm to
determine whether these constraints are active at the solution. In the case of active-set
methods and gradient projection methods (described below), this indecisiveness can cause
the algorithm to zigzag as the iterates move on and off the weakly active constraints on
successive iterations. Safeguards must be used to prevent such behavior.
16.5
ACTIVE-SET METHODS FOR CONVEX QPs
We now describe active-set methods for solving quadratic programs of the form (16.1)
containing equality and inequality constraints. We consider only the convex case, in which
the matrix G in (16.1a) is positive semidefinite. The case in which G is an indefinite matrix
raises complications in the algorithms and is outside the scope of this book. We refer to
Gould [147] for a discussion of nonconvex QPs.
If the contents of the optimal active set (16.36) were known in advance, we could
find the solution x ∗ by applying one of the techniques for equality-constrained QP of
Sections 16.2 and 16.3 to the problem
min q(x) 12 x T Gx + x T c
x
subject to
aiT x bi , i ∈ A(x ∗ ).
Of course, we usually do not have prior knowledge of A(x ∗ ) and, as we now see, determination of this set is the main challenge facing algorithms for inequality-constrained
QP.
We have already encountered an active-set approach for linear programming in Chapter 13, namely, the simplex method. In essence, the simplex method starts by making a guess
of the optimal active set, then repeatedly uses gradient and Lagrange multiplier information
to drop one index from the current estimate of A(x ∗ ) and add a new index, until optimality
467
468
CHAPTER 16.
QUADRATIC PROGRAMMING
is detected. Active-set methods for QP differ from the simplex method in that the iterates
(and the solution x ∗ ) are not necessarily vertices of the feasible region.
Active-set methods for QP come in three varieties: primal, dual, and primal-dual. We
restrict our discussion to primal methods, which generate iterates that remain feasible with
respect to the primal problem (16.1) while steadily decreasing the objective function q(x).
Primal active-set methods find a step from one iterate to the next by solving a quadratic
subproblem in which some of the inequality constraints (16.1c), and all the equality constraints (16.1b), are imposed as equalities. This subset is referred to as the working set and
is denoted at the kth iterate xk by Wk . An important requirement we impose on Wk is that
the gradients ai of the constraints in the working set be linearly independent, even when
the full set of active constraints at that point has linearly dependent gradients.
Given an iterate xk and the working set Wk , we first check whether xk minimizes
the quadratic q in the subspace defined by the working set. If not, we compute a step p
by solving an equality-constrained QP subproblem in which the constraints corresponding
to the working set Wk are regarded as equalities and all other constraints are temporarily
disregarded. To express this subproblem in terms of the step p, we define
p x − xk ,
gk Gxk + c.
By substituting for x into the objective function (16.1a), we find that
q(x) q(xk + p) 12 p T Gp + gkT p + ρk ,
where ρk 12 xkT Gxk + c T xk is independent of p. Since we can drop ρk from the objective
without changing the solution of the problem, we can write the QP subproblem to be solved
at the kth iteration as follows:
min
p
subject to
1
2
p T Gp + gkT p
aiT p 0,
i ∈ Wk .
(16.39a)
(16.39b)
We denote the solution of this subproblem by pk . Note that for each i ∈ Wk , the value of
aiT x does not change as we move along pk , since we have aiT (xk + αpk ) aiT xk bi for
all α. Since the constraints in Wk were satisfied at xk , they are also satisfied at xk + αpk , for
any value of α. Since G is positive definite, the solution of (16.39) can be computed by any
of the techniques described in Section 16.2.
Supposing for the moment that the optimal pk from (16.39) is nonzero, we need to
decide how far to move along this direction. If xk + pk is feasible with respect to all the
constraints, we set xk+1 xk + pk . Otherwise, we set
xk+1 xk + αk pk ,
(16.40)
16.5.
ACTIVE-SET METHODS
FOR
CONVEX QPS
where the step-length parameter αk is chosen to be the largest value in the range [0, 1] for
which all constraints are satisfied. We can derive an explicit definition of αk by considering
what happens to the constraints i ∈
/ Wk , since the constraints i ∈ Wk will certainly be
/ Wk , then for all αk ≥ 0 we
satisfied regardless of the choice of αk . If aiT pk ≥ 0 for some i ∈
have aiT (xk + αk pk ) ≥ aiT xk ≥ bi . Hence, constraint i will be satisfied for all nonnegative
/ Wk , however, we
choices of the step-length parameter. Whenever aiT pk < 0 for some i ∈
have that aiT (xk + αk pk ) ≥ bi only if
αk ≤
bi − aiT xk
.
aiT pk
To maximize the decrease in q, we want αk to be as large as possible in [0, 1] subject to
retaining feasibility, so we obtain the following definition:
def
αk min
bi − aiT xk
1,
min
aiT pk
i ∈W
/ k , aiT pk <0
.
(16.41)
We call the constraints i for which the minimum in (16.41) is achieved the blocking constraints. (If αk 1 and no new constraints are active at xk + αk pk , then there are no blocking
constraints on this iteration.) Note that it is quite possible for αk to be zero, because we
could have aiT pk < 0 for some constraint i that is active at xk but not a member of the
current working set Wk .
If αk < 1, that is, the step along pk was blocked by some constraint not in Wk , a new
working set Wk+1 is constructed by adding one of the blocking constraints to Wk .
We continue to iterate in this manner, adding constraints to the working set until we
reach a point x̂ that minimizes the quadratic objective function over its current working set
Ŵ. It is easy to recognize such a point because the subproblem (16.39) has solution p 0.
Since p 0 satisfies the optimality conditions (16.5) for (16.39), we have that
ai λ̂i g G x̂ + c,
(16.42)
i∈Ŵ
for some Lagrange multipliers λ̂i , i ∈ Ŵ. It follows that x̂ and λ̂ satisfy the first KKT
condition (16.37a), if we define the multipliers corresponding to the inequality constraints
that are not in the working set to be zero. Because of the control imposed on the step length,
x̂ is also feasible with respect to all the constraints, so the second and third KKT conditions
(16.37b) and (16.37c) are satisfied at this point.
We now examine the signs of the multipliers corresponding to the inequality constraints in the working set, that is, the indices i ∈ Ŵ ∩ I. If these multipliers are all
nonnegative, the fourth KKT condition (16.37d) is also satisfied, so we conclude that x̂ is a
KKT point for the original problem (16.1). In fact, since G is positive semidefinite, we have
469
470
CHAPTER 16.
QUADRATIC PROGRAMMING
from Theorem 16.4 that x̂ is a global solution of (16.1). (As noted after Theorem 16.4, x̂ is
a strict local minimizer and the unique global solution if G is positive definite.)
If, on the other hand, one or more of the multipliers λ̂ j , j ∈ Ŵ ∩ I, is negative,
the condition (16.37d) is not satisfied and the objective function q(·) may be decreased
by dropping one of these constraints, as shown in Section 12.3. Thus, we remove an index
j corresponding to one of the negative multipliers from the working set and solve a new
subproblem (16.39) for the new step. We show in the following theorem that this strategy
produces a direction p at the next iteration that is feasible with respect to the dropped
constraint. We continue to assume that the constraint gradients ai for i in the working
set are linearly independent. After the algorithm has been fully stated, we discuss how this
property can be maintained.
Theorem 16.5.
Suppose that the point x̂ satisfies first-order conditions for the equality-constrained
subproblem with working set Ŵ; that is, equation (16.42) is satisfied along with aiT x̂ bi for
all i ∈ Ŵ. Suppose, too, that the constraint gradients ai , i ∈ Ŵ, are linearly independent and
that there is an index j ∈ Ŵ such that λ̂ j < 0. Let p be the solution obtained by dropping the
constraint j and solving the following subproblem:
min
p
1
2
p T Gp + (G x̂ + c)T p,
subject to aiT p 0, for all i ∈ Ŵ with i j.
(16.43a)
(16.43b)
Then p is a feasible direction for constraint j, that is, a Tj p ≥ 0. Moreover, if p satisfies secondorder sufficient conditions for (16.43), then we have that a Tj p > 0, and that p is a descent
direction for q(·).
PROOF. Since p solves (16.43), we have from the results of Section 16.1 that there are
multipliers λ̃i , for all i ∈ Ŵ with i j, such that
λ̃i ai Gp + (G x̂ + c).
(16.44)
i∈Ŵ, i j
In addition, we have by second-order necessary conditions that if Z is a null-space basis
vector for the matrix
aiT
i∈Ŵ, i j
,
then Z T G Z is positive semidefinite. Clearly, p has the form p Z pZ for some vector pZ ,
so it follows that p T Gp ≥ 0.
16.5.
ACTIVE-SET METHODS
FOR
CONVEX QPS
We have made the assumption that x̂ and Ŵ satisfy the relation (16.42). By subtracting
(16.42) from (16.44), we obtain
(λ̃i − λ̂i )ai − λ̂ j a j Gp.
(16.45)
i∈Ŵ, i j
By taking inner products of both sides with p and using the fact that aiT p 0 for all i ∈ Ŵ
with i j, we have that
− λ̂ j a Tj p p T Gp.
(16.46)
Since p T Gp ≥ 0 and λ̂ j < 0 by assumption, it follows that a Tj p ≥ 0.
If the second-order sufficient conditions of Section 12.5 are satisfied, we have that
Z T G Z defined above is positive definite. From (16.46), we can have a Tj p 0 only if
p T Gp pZT Z T G Z pZ 0, which happens only if pZ 0 and p 0. But if p 0, then
by substituting into (16.45) and using linear independence of ai for i ∈ Ŵ, we must have
that λ̂ j 0, which contradicts our choice of j. We conclude that p T Gp > 0 in (16.46), and
therefore a Tj p > 0 whenever p satisfies the second-order sufficient conditions for (16.43).
The claim that p is a descent direction for q(·) is proved in Theorem 16.6 below.
While any index j for which λ̂ j < 0 usually will yield a direction p along which the
algorithm can make progress, the most negative multiplier is often chosen in practice (and
in the algorithm specified below). This choice is motivated by the sensitivity analysis given
in Chapter 12, which shows that the rate of decrease in the objective function when one
constraint is removed is proportional to the magnitude of the Lagrange multiplier for that
constraint. As in linear programming, however, the step along the resulting direction may
be short (as when it is blocked by a new constraint), so the amount of decrease in q is not
guaranteed to be greater than for other possible choices of j.
We conclude with a result that shows that whenever pk obtained from (16.39) is
nonzero and satisfies second-order sufficient optimality conditions for the current working
set, it is a direction of strict descent for q(·).
Theorem 16.6.
Suppose that the solution pk of (16.39) is nonzero and satisfies the second-order sufficient
conditions for optimality for that problem. Then the function q(·) is strictly decreasing along
the direction pk .
PROOF. Since pk satisfies the second-order conditions, that is, Z T G Z is positive definite
for the matrix Z whose columns are a basis of the null space of the constraints (16.39b), we
have by applying Theorem 16.2 to (16.39) that pk is the unique global solution of (16.39).
Since p 0 is also a feasible point for (16.39), its objective value in (16.39a) must be larger
471
472
CHAPTER 16.
QUADRATIC PROGRAMMING
than that of pk , so we have
1
2
pkT Gpk + gkT pk < 0.
Since pkT Gpk ≥ 0 by convexity, this inequality implies that gkT pk < 0. Therefore, we have
q(xk + αk pk ) q(xk ) + αgkT pk + 12 α 2 pkT Gpk < q(xk ),
for all α > 0 sufficiently small.
When G is positive definite—the strictly convex case—the second-order sufficient
conditions are satisfied for all feasible subproblems of the form (16.39). Hence, it follows
from the result above that we obtain a strict decrease in q(·) whenever pk 0. This fact is
significant when we discuss finite termination of the algorithm.
SPECIFICATION OF THE ACTIVE-SET METHOD FOR CONVEX QP
Having described the active-set algorithm for convex QP, we now present the following
formal specification. We assume that the objective function q is bounded in the feasible set
(16.1b), (16.1c).
Algorithm 16.3 (Active-Set Method for Convex QP).
Compute a feasible starting point x0 ;
Set W0 to be a subset of the active constraints at x0 ;
for k 0, 1, 2, . . .
Solve (16.39) to find pk ;
if pk 0
Compute Lagrange multipliers λ̂i that satisfy (16.42),
with Ŵ Wk ;
if λ̂i ≥ 0 for all i ∈ Wk ∩ I
stop with solution x ∗ xk ;
else
j ← arg min j∈Wk ∩I λ̂ j ;
xk+1 ← xk ; Wk+1 ← Wk \{ j};
else (* pk 0 *)
Compute αk from (16.41);
xk+1 ← xk + αk pk ;
if there are blocking constraints
Obtain Wk+1 by adding one of the blocking
constraints to Wk ;
else
Wk+1 ← Wk ;
end (for)
16.5.
ACTIVE-SET METHODS
FOR
CONVEX QPS
Various techniques can be used to determine an initial feasible point. One such is
to use the “Phase I” approach for linear programming described in Chapter 13. Though
no significant modifications are needed to generalize this method from linear programming to quadratic programming, we describe a variant here that allows the user to
supply an initial estimate x̃ of the vector x. This estimate need not be feasible, but a
good choice based on knowledge of the QP may reduce the work needed in the Phase I
step.
Given x̃, we define the following feasibility linear program:
min e T z
(x,z)
subject to aiT x + γi z i bi ,
i ∈ E,
+ γi z i ≥ bi ,
i ∈ I,
aiT x
z ≥ 0,
where e (1, 1, . . . , 1)T , γi −sign(aiT x̃ − bi ) for i ∈ E, and γi 1 for i ∈ I. A feasible
initial point for this problem is then
x x̃,
z i |aiT x̃ − bi | (i ∈ E),
z i max(bi − aiT x̃, 0) (i ∈ I).
It is easy to verify that if x̃ is feasible for the original problem (16.1), then ( x̃, 0) is optimal for
the feasibility subproblem. In general, if the original problem has feasible points, then the
optimal objective value in the subproblem is zero, and any solution of the subproblem yields
a feasible point for the original problem. The initial working set W0 for Algorithm 16.3 can
be found by taking a linearly independent subset of the active constraints at the solution of
the feasibility problem.
An alternative approach is a penalty (or “big M”) method, which does away with the
“Phase I” and instead includes a measure of infeasibility in the objective that is guaranteed
to be zero at the solution. That is, we introduce a scalar artificial variable η into (16.1) to
measure the constraint violation, and we solve the problem
min 12 x T Gx + x T c + Mη,
(x,η)
subject to
(aiT x − bi ) ≤ η,
i ∈ E,
− bi ) ≤ η,
i ∈ E,
−(aiT x
bi −
aiT x
≤ η,
(16.47)
i ∈ I,
0 ≤ η,
for some large positive value of M. It can be shown by applying the theory of exact penalty
functions (see Chapter 17) that whenever there exist feasible points for the original problem
(16.1), then for all M sufficiently large, the solution of (16.47) will have η 0, with an x
component that is a solution for (16.1).
473
474
CHAPTER 16.
QUADRATIC PROGRAMMING
Our strategy is to use some heuristic to choose a value of M and solve (16.47) by
the usual means. If the solution we obtain has a positive value of η, we increase M and try
again. Note that a feasible point is easy to obtain for the subproblem (16.47): We set x x̃
(where, as before, x̃ is the user-supplied initial guess) and choose η large enough that all the
constraints in (16.47) are satisfied. This approach is, in fact, an exact penalty method using
the ∞ norm; see Chapter 17.
A variant of (16.47) that penalizes the 1 norm of the constraint violation rather than
the ∞ norm is as follows:
min 1 x T Gx
(x,s,t,v) 2
+ x T c + MeET (s + t) + MeIT v
subject to aiT x − bi + si − ti 0, i ∈ E,
aiT x − bi + vi ≥ 0, i ∈ I,
(16.48)
s ≥ 0, t ≥ 0, v ≥ 0.
Here, eE is the vector (1, 1, . . . , 1)T of length |E|; similarly for eI . The slack variables si , ti ,
and vi soak up any infeasibility in the constraints.
In the following example we use subscripts on the vectors x and p to denote their
components, and we use superscripts to indicate the iteration index. For example, x1 denotes
the first component, while x 4 denotes the fourth iterate of the vector x.
x2
(2,2)
x4
x5
(4,1)
(0,1)
x 2, x 3
Figure 16.3
(2,0)
x 0, x 1
Iterates of the active-set method.
x1
16.5.
ACTIVE-SET METHODS
FOR
CONVEX QPS
❏ EXAMPLE 16.4
We apply Algorithm 16.3 to the following simple 2-dimensional problem illustrated
in Figure 16.3.
min q(x) (x1 − 1)2 + (x2 − 2.5)2
(16.49a)
x
subject to
x1 − 2x2 + 2 ≥ 0,
(16.49b)
−x1 − 2x2 + 6 ≥ 0,
(16.49c)
−x1 + 2x2 + 2 ≥ 0,
(16.49d)
x1 ≥ 0,
(16.49e)
x2 ≥ 0.
(16.49f)
We refer the constraints, in order, by indices 1 through 5. For this problem it is easy to
determine a feasible initial point; say x 0 (2, 0)T . Constraints 3 and 5 are active at this
point, and we set W0 {3, 5}. (Note that we could just as validly have chosen W0 {5}
or W0 {3} or even W ∅; each choice would lead the algorithm to perform somewhat
differently.)
Since x 0 lies on a vertex of the feasible region, it is obviously a minimizer of the
objective function q with respect to the working set W0 ; that is, the solution of (16.39) with
k 0 is p 0. We can then use (16.42) to find the multipliers λ̂3 and λ̂5 associated with
the active constraints. Substitution of the data from our problem into (16.42) yields
−1
2
λ̂3 +
0
1
λ̂5 2
−5
,
which has the solution (λ̂3 , λ̂5 ) (−2, −1).
We now remove constraint 3 from the working set, because it has the most negative
multiplier, and set W1 {5}. We begin iteration 1 by finding the solution of (16.39) for
k 1, which is p 1 (−1, 0)T . The step-length formula (16.41) yields α1 1, and the new
iterate is x 2 (1, 0)T .
There are no blocking constraints, so that W2 W1 {5}, and we find at the start of
iteration 2 that the solution of (16.39) is p 2 0. From (16.42) we deduce that the Lagrange
multiplier for the lone working constraint is λ̂5 −5, so we drop 5 from the working set to
obtain W3 ∅.
Iteration 3 starts by solving the unconstrained problem, to obtain the solution p 3 (0, 2.5)T . The formula (16.41) yields a step length of α3 0.6 and a new iterate x 4 (1, 1.5)T . There is a single blocking constraint (constraint 1), so we obtain W4 {1}. The
solution of (16.39) for k 4 is then p 4 (0.4, 0.2)T , and the new step length is 1. There
are no blocking constraints on this step, so the next working set is unchanged: W5 {1}.
The new iterate is x 5 (1.4, 1.7)T .
475
476
CHAPTER 16.
QUADRATIC PROGRAMMING
Finally, we solve (16.39) for k 5 to obtain a solution p 5 0. The formula (16.42)
yields a multiplier λ̂1 0.8, so we have found the solution. We set x ∗ (1.4, 1.7)T and
terminate.
❐
FURTHER REMARKS ON THE ACTIVE-SET METHOD
We noted above that there is flexibility in the choice of the initial working set and that
each initial choice leads to a different iteration sequence. When the initial active constraints
have independent gradients, as above, we can include them all in W0 . Alternatively, we can
select a subset. For instance, if in the example above we have chosen W0 {3}, the first
iterate would have yielded p 0 (0.2, 0.1)T and a new iterate of x 1 (2.2, 0.1)T . If we
had chosen W0 {5}, we would have moved immediately to the new iterate x 1 (1, 0)T ,
without first performing the operation of dropping the index 3, as is done in the example. If
we had selected W0 ∅, we would have obtained p 1 (−1, 2.5)T , α1 23 , a new iterate of
x 1 ( 34 , 53 )T , and a new working set of W1 {1}. The solution x ∗ would have been found
on the next iteration.
Even if the initial working set W0 coincides with the initial active set, the sets Wk
and A(x k ) may differ at later iterations. For instance, when a particular step encounters
more than one blocking constraint, just one of them is added to the working set, so the
identification between Wk and A(x k ) is broken. Moreover, subsequent iterates differ in
general according to what choice is made.
We require the constraint gradients in W0 to be linearly independent, and our strategy
for modifying the working set ensures that this same property holds for all subsequent
working sets Wk . When we encounter a blocking constraint on a particular step, its constraint
normal cannot be a linear combination of the normals ai in the current working set (see
Exercise 16.18). Hence, linear independence is maintained after the blocking constraint is
added to the working set. On the other hand, deletion of an index from the working set
cannot introduce linear dependence.
The strategy of removing the constraint corresponding to the most negative Lagrange
multiplier often works well in practice but has the disadvantage that it is susceptible to the
scaling of the constraints. (By multiplying constraint i by some factor β > 0 we do not
change the geometry of the optimization problem, but we introduce a scaling of 1/β to
the corresponding multiplier λi .) Choice of the most negative multiplier is analogous to
Dantzig’s original pivot rule for the simplex method in linear programming (see Chapter 13)
and, as we noted there, strategies that are less sensitive to scaling often give better results.
We do not discuss this advanced topic further.
We note that the strategy of adding or deleting at most one constraint at each iteration
of the Algorithm 16.3 places a natural lower bound on the number of iterations needed
to reach optimality. Suppose, for instance, that we have a problem in which m inequality
constraints are active at the solution x ∗ but that we start from a point x 0 that is strictly
16.5.
ACTIVE-SET METHODS
FOR
CONVEX QPS
feasible with respect to all the inequality constraints. In this case, the algorithm will need
at least m iterations to move from x 0 to x ∗ . Even more iterations will be required if the
algorithm adds some constraint j to the working set at some iteration, only to remove it at
a later step.
FINITE TERMINATION OF ACTIVE-SET ALGORITHM ON STRICTLY CONVEX
QPs
It is not difficult to show that, under certain assumptions, Algorithm 16.3 converges
for strictly convex QPs, that is, it identifies the solution x ∗ in a finite number of iterations.
This claim is certainly true if we assume that the method always takes a nonzero step length
αk whenever the direction pk computed from (16.39) is nonzero. Our argument proceeds
as follows:
• If the solution of (16.39) is pk 0, the current point xk is the unique global minimizer
of q(·) for the working set Wk ; see Theorem 16.6. If it is not the solution of the
original problem (16.1) (that is, at least one of the Lagrange multipliers is negative),
Theorems 16.5 and 16.6 together show that the step pk+1 computed after a constraint
is dropped will be a strict decrease direction for q(·). Therefore, because of our
assumption αk > 0, we have that the value of q is lower than q(xk ) at all subsequent
iterations. It follows that the algorithm can never return to the working set Wk ,
because subsequent iterates have values of q that are lower than the global minimizer
for this working set.
• The algorithm encounters an iterate k for which pk 0 solves (16.39) at least on
every nth iteration. To demonstrate this claim, we note that for any k at which pk 0,
either we have αk 1 (in which case we reach the minimizer of q on the current
working set Wk , so that the next iteration will yield pk+1 0), or else a constraint
is added to the working set Wk . If the latter situation occurs repeatedly, then after
at most n iterations the working set will contain n indices, which correspond to n
linearly independent vectors. The solution of (16.39) will then be pk 0, since only
the zero vector will satisfy the constraints (16.39b).
• Taken together, the two statements above indicate that the algorithm finds the global
minimum of q on its current working set periodically (at least once every n iterations)
and that, having done so, it never visits this particular working set again. It follows
that, since there are only a finite number of possible working sets, the algorithm
cannot iterate forever. Eventually, it encounters a minimizer for a current working set
that satisfies optimality conditions for (16.1), and it terminates with a solution.
The assumption that we can always take a nonzero step along a nonzero descent
direction pk calculated from (16.39) guarantees that the algorithm does not undergo cycling.
This term refers to the situation in which a sequence of consecutive iterations results in no
movement in iterate x, while the working set Wk undergoes deletions and additions of indices
477
478
CHAPTER 16.
QUADRATIC PROGRAMMING
and eventually repeats itself. That is, for some integers k and l ≥ 1, we have that x k x k+l
and Wk Wk+l . At each iterate in the cycle, a constraint is dropped (as in Theorem 16.5),
but a new constraint i ∈
/ Wk is encountered immediately without any movement along
the computed direction p. Procedures for handling degeneracy and cycling in quadratic
programming are similar to those for linear programming discussed in Chapter 13; we do
not discuss them here. Most QP implementations simply ignore the possibility of cycling.
UPDATING FACTORIZATIONS
We have seen that the step computation in the active-set method given in Algorithm 16.3 requires the solution of the equality-constrained subproblem (16.39). As
mentioned at the beginning of this chapter, this computation amounts to solving the KKT
system (16.5). Since the working set can change by just one index at every iteration, the KKT
matrix differs in at most one row and one column from the previous iteration’s KKT matrix.
Indeed, G remains fixed, whereas the matrix A of constraint gradients corresponding to the
current working set may change through addition and/or deletion of a single row.
It follows from this observation that we can compute the matrix factors needed to solve
(16.39) at the current iteration by updating the factors computed at the previous iteration,
rather than recomputing them from scratch. These updating techniques are crucial to the
efficiency of active-set methods.
We limit our discussion to the case in which the step is computed with the null-space
method (16.17)–(16.20). Suppose that A has m linearly independent rows and assume that
the bases Y and Z are defined by means of a QR factorization of A (see Section 15.3 for
details). Thus
A Q
T
R
0
Q1
Q2
R
(16.50)
0
(see (15.21)), where is a permutation matrix; R is square, upper triangular and nonsingular; Q Q 1 Q 2 is n × n orthogonal; and Q 1 and R both have m columns while Q 2
has n − m columns. As noted in Chapter 15, we can choose Z to be simply the orthonormal
matrix Q 2 .
Suppose that one constraint is added to the working set at the next iteration, so that
the new constraint matrix is Ā T A T a , where a is a column vector of length n
such that Ā T retains full column rank. As we now show, there is an economical way to
update the Q and R factors in (16.50) to obtain new factors (and hence a new null-space
basis matrix Z̄ , with n − m − 1 columns) for the expanded matrix Ā. Note first that, since
Q 1 Q 1T + Q 2 Q 2T I , we have
Ā
T
0
0
1
A T
a
Q
R
Q 1T a
0
Q 2T a
.
(16.51)
16.5.
ACTIVE-SET METHODS
CONVEX QPS
FOR
We can now define an orthogonal matrix Q̂ that transforms the vector Q 2T a to a vector in
which all elements except the first are zero. That is, we have
Q̂(Q 2T a)
γ
,
0
where γ is a scalar. (Since Q̂ is orthogonal, we have Q 2T a |γ |.) From (16.51) we now
have
Ā T
0
0
1
⎡
⎢
Q⎣
⎤
Q 1T a
⎥
I
γ
⎦Q
T
Q̂
0
0
R
0
0
Q̂ T
⎡
R
⎢
⎣ 0
Q 1T a
γ
0
⎤
⎥
⎦.
0
This factorization has the form
¯ Q̄
Ā T where
¯ 0
0
1
, Q̄ Q
I
0
0
Q̂ T
R̄
,
0
)
Q 2 Q̂
Q1
T
*
, R̄ R
Q 1T a
0
γ
.
We can therefore choose Z̄ to be the last n −m −1 columns of Q 2 Q̂ T . If we know Z explicitly
and need an explicit representation of Z̄ , we need to account for the cost of obtaining Q̂ and
the cost of forming the product Q 2 Q̂ T Z Q̂ T . Because of the special structure of Q̂, this
cost is of order n(n − m), compared to the cost of computing (16.50) from scratch, which is
of order n 2 m. The updating strategy is less expensive, especially when the null space is small
(that is, when n − m n).
An updating technique can also be designed for the case in which a row is removed
from A. This operation has the effect of deleting a column from R in (16.50), thus disturbing
the upper triangular property of this matrix by introducing a number of nonzeros on the
diagonal immediately below the main diagonal of the matrix. Upper triangularity can be
restored by applying a sequence of plane rotations. These rotations introduce a number
of inexpensive transformations into the first m columns of Q, and the updated null-space
matrix is obtained by selecting the last n − m + 1 columns from this matrix after the
transformations are complete. The new null-space basis in this case has the form
Z̄ z̄
Z
,
(16.52)
that is, the current matrix Z is augmented by a single column. The total cost of this
operation varies with the location of the removed column in A but is in general cheaper
479
480
CHAPTER 16.
QUADRATIC PROGRAMMING
than recomputing a QR factorization from scratch. For details of these procedures, see Gill
et al. [124, Section 5].
We now consider the reduced Hessian. Because of the special form of (16.39), we have
h 0 in (16.5), and the step pY given in (16.18) is zero. Thus from (16.19), the null-space
component pZ is the solution of
(Z T G Z ) pZ −Z T g.
(16.53)
We can sometimes find ways of updating the factorization of the reduced Hessian Z T G Z
after Z has changed. Suppose that we have the Cholesky factorization of the current reduced
Hessian, written as
ZT G Z L LT ,
and that at the next step Z changes as in (16.52), gaining a column after deletion of a
constraint. A series of inexpensive, elementary operations can be used to transform the
Cholesky factor L into the new factor L̄ for the new reduced Hessian Z̄ T G Z̄ .
A variety of other simplifications are possible. For example, as discussed in Section 16.7, we can update the reduced gradient Z T g at the same time as we update Z to
Z̄ .
16.6
INTERIOR-POINT METHODS
The interior-point approach can be applied to convex quadratic programs through a simple
extension of the linear-programming algorithms described in Chapter 14. The resulting
primal-dual algorithms are easy to describe and are quite efficient on many types of problems.
Extensions of interior-point methods to nonconvex problems are discussed in Chapter 19.
For simplicity, we restrict our attention to convex quadratic programs with inequality
constraints, which we write as follows:
min
x
subject to
q(x) 12 x T Gx + x T c
(16.54a)
Ax ≥ b,
(16.54b)
where G is symmetric and positive semidefinite and where the m×n matrix A and right-hand
side b are defined by
A [ai ]i∈I ,
b [bi ]i∈I ,
I {1, 2, . . . , m}.
(If equality constraints are also present, they can be accommodated with simple extensions
to the approaches described below.) Rewriting the KKT conditions (16.37) in this notation,
16.6.
INTERIOR-POINT METHODS
we obtain
Gx − A T λ + c 0,
Ax − b ≥ 0,
(Ax − b)i λi 0,
i 1, 2, . . . , m,
λ ≥ 0.
By introducing the slack vector y ≥ 0, we can rewrite these conditions as
Gx − A T λ + c 0,
(16.55a)
Ax − y − b 0,
(16.55b)
yi λi 0,
i 1, 2, . . . , m,
(16.55c)
(y, λ) ≥ 0.
(16.55d)
Since we assume that G is positive semidefinite, these KKT conditions are not only necessary
but also sufficient (see Theorem 16.4), so we can solve the convex quadratic program (16.54)
by finding solutions of the system (16.55).
Given a current iterate (x, y, λ) that satisfies (y, λ) > 0, we can define a
complementarity measure µ by
µ
yT λ
.
m
(16.56)
As in Chapter 14, we derive path-following, primal-dual methods by considering the
perturbed KKT conditions given by
⎡
⎢
F(x, y, λ; σ µ) ⎣
Gx − A T λ + c
Ax − y − b
⎤
⎥
⎦ 0,
(16.57)
Ye − σ µe
where
Y diag(y1 , y2 , . . . , ym ),
diag(λ1 , λ2 , . . . , λm ),
e (1, 1, . . . , 1)T ,
and σ ∈ [0, 1]. The solutions of (16.57) for all positive values of σ and µ define the central
path, which is a trajectory that leads to the solution of the quadratic program as σ µ tends
to zero.
By fixing µ and applying Newton’s method to (16.57), we obtain the linear system
⎡
⎢
⎣ A
0
−A T
−I
0
0
Y
G
⎤⎡
x
⎤
⎡
⎥⎢
⎥ ⎢
⎦ ⎣ y ⎦ ⎣
λ
−rd
−r p
−Ye + σ µe
⎤
⎥
⎦,
(16.58)
481
482
CHAPTER 16.
QUADRATIC PROGRAMMING
where
rd Gx − A T λ + c,
r p Ax − y − b.
(16.59)
(x + , y + , λ+ ) (x, y, λ) + α(x, y, λ),
(16.60)
We obtain the next iterate by setting
where α is chosen to retain the inequality (y + , λ+ ) > 0 and possibly to satisfy various other
conditions.
In the rest of the chapter we discuss several enhancements of this primal-dual iteration
that make it effective in practice.
SOLVING THE PRIMAL-DUAL SYSTEM
The major computational operation in the interior-point method is the solution of
the system (16.58). The coefficient matrix in this system can be much more costly to factor
than the matrix (14.9) arising in linear programming because of the presence of the Hessian
matrix G. It is therefore important to exploit the structure of (16.58) by choosing a suitable
direct factorization algorithm, or by choosing an appropriate preconditioner for an iterative
solver.
As in Chapter 14, the system (16.58) may be restated in more compact forms. The
“augmented system” form is
G
−A T
A
−1 Y
x
λ
−rd
−r p + (−y + σ µ−1 e)
.
(16.61)
After a simple transformation to symmetric form, a symmetric indefinite factorization
scheme can be applied to the coefficient matrix in this system. The “normal equations” form
(14.44a) is
(G + A T Y −1 A)x −rd + A T Y −1 [−r p − y + σ µ−1 e],
(16.62)
which can be solved by means of a modified Cholesky algorithm. This approach is effective
if the term A T (Y −1 )A is not too dense compared with G, and it has the advantage of
being much smaller than (16.61) if there are many inequality constraints.
The projected CG method of Algorithm 16.2 can also be effective for solving the
primal-dual system. We can rewrite (16.58) in the form
⎡
G
⎢
⎣ 0
A
0
−A T
Y −1 I
−I
0
⎤⎡
x
⎤
⎡
−rd
⎤
⎥⎢
⎥ ⎢
⎥
⎦ ⎣ y ⎦ ⎣ −e + σ µY −1 e ⎦ ,
λ
−r p
(16.63)
16.6.
INTERIOR-POINT METHODS
and observe that these are the optimality conditions for an equality-constrained convex
quadratic program of the form (16.3), in which the variable is (x, y). Hence, we can
make appropriate substitutions and solve this system using Algorithm 16.2. This approach
may be useful for problems in which the direct factorization cannot be performed due to
excessive memory demands. The projected CG method does not require that the matrix G
be formed or factored; it requires only matrix-vector products.
STEP LENGTH SELECTION
We mentioned in Chapter 14 that interior-point methods for linear programming are
more efficient if different step lengths α pri , α dual are used for the primal and dual variables.
Equation (14.37) indicates that the greatest reduction in the residuals rb and rc is obtained
by choosing the largest admissible primal and dual step lengths. The situation is different in
quadratic programming. Suppose that we define the new iterate as
(x + , y + ) (x, y) + α pri (x, y),
λ+ λ + α dual λ,
(16.64)
where α pri and α dual are step lengths that ensure the positivity of (y + , λ+ ). By using (16.58)
and (16.59), we see that the new residuals satisfy the following relations:
pri
r+
p (1 − α )r p ,
rd+
(1 − α
dual
)rd + (α
(16.65a)
pri
−α
dual
)Gx.
(16.65b)
If α pri α dual α then both residuals decrease linearly for all α ∈ (0, 1). For different step
lengths, however, the dual residual rd+ may increase for certain choices of α pri , α dual , possibly
causing divergence of the interior-point iteration.
pri
One option is to use equal step lengths, as in (16.60), and to set α min(ατ , ατdual ),
where
ατpri max{α ∈ (0, 1] : y + αy ≥ (1 − τ )y},
(16.66a)
max{α ∈ (0, 1] : λ + αλ ≥ (1 − τ )λ};
(16.66b)
ατdual
the parameter τ ∈ (0, 1) controls how far we back off from the maximum step for which the
conditions y + αy ≥ 0 and λ + αλ ≥ 0 are satisfied. Numerical experience has shown,
however, that using different step lengths in the primal and dual variables often leads to
faster convergence. One way to choose unequal step lengths is to select (α pri , α dual ) so as to
(approximately) minimize the optimality measure
Gx + − A T λ+ + c22 + Ax + − y + − b22 + (y + )T z + ,
subject to 0 ≤ α pri ≤ ατ and 0 ≤ α dual ≤ ατdual , where x + , y + , λ+ are defined as a function
of the step lengths through (16.64).
pri
483
484
CHAPTER 16.
QUADRATIC PROGRAMMING
A PRACTICAL PRIMAL-DUAL METHOD
The most popular interior-point method for convex QP is based on Mehrotra’s
predictor-corrector, originally developed for linear programming (see Section 14.2). The
extension to quadratic programming is straightforward, as we now show.
First, we compute an affine scaling step (x aff , y aff , λaff ) by setting σ 0 in
(16.58). We improve upon this step by computing a corrector step, which is defined following
the same reasoning that leads to (14.31). Next, we compute the centering parameter σ using
(14.34). The total step is obtained by solving the following system (cf. (14.35)):
⎡
⎢
⎣ A
0
−A T
−I
0
0
Y
G
⎤⎡
x
⎤
⎡
⎥ ⎢
⎥⎢
⎦ ⎣ y ⎦ ⎣
λ
⎤
−rd
⎥
⎦.
−r p
−Ye − Y
aff
aff
(16.67)
e + σ µe
We now specify the algorithm. For simplicity, we will assume in our description that
equal step lengths are used in the primal and dual variables though, as noted above, unequal
step lengths can give slightly faster convergence.
Algorithm 16.4 (Predictor-Corrector Algorithm for QP).
Compute (x0 , y0 , λ0 ) with (y0 , λ0 ) > 0;
for k 0, 1, 2, . . .
Set (x, y, λ) (xk , yk , λk ) and solve (16.58) with σ 0 for
(x aff , y aff , λaff );
Calculate µ y T λ/m;
Calculate α̂aff max{α ∈ (0, 1] | (y, λ) + α(y aff , λaff ) ≥ 0};
Calculate µaff (y + α̂ aff y aff )T (λ + α̂ aff λaff )/m;
Set centering parameter to σ (µaff /µ)3 ;
Solve (16.67) for (x, y, λ);
pri
) (see (16.66));
Choose τk ∈ (0, 1) and set α̂ min(ατk , ατdual
k
Set (xk+1 , yk+1 , λk+1 ) (xk , yk , λk ) + α̂(x, y, λ);
end (for)
We can choose τk to approach 1 as the iterates approach the solution, to accelerate the
convergence.
As for linear programming, efficiency and robustness of this approach is greatly
enhanced if we choose a good starting point. This selection can be done in several ways. The
following simple heuristic accepts an initial point (x̄, ȳ, λ̄) from the user and moves it far
enough away from the boundary of the region (y, λ) ≥ 0 to permit the algorithm to take
long steps on early iterations. First, we compute the affine scaling step (x aff , y aff , λaff )
16.7.
THE GRADIENT PROJECTION METHOD
from the user-supplied initial point ( x̄, ȳ, λ̄), then set
y0 max(1, | ȳ + y aff |),
λ0 max(1, |λ̄ + λaff |),
x0 x̄,
where the max and absolute values are applied component-wise.
We conclude this section by contrasting some of the properties of active-set and
interior-point methods for convex quadratic programming. Active-set methods generally
require a large number of steps in which each search direction is relatively inexpensive to
compute, while interior-point methods take a smaller number of more expensive steps.
Active-set methods are more complicated to implement, particularly if the procedures for
updating matrix factorizations try to take advantage of sparsity or structure in G and A. By
contrast, the nonzero structure of the matrix to be factored at each interior-point iteration
remains the same at all iterations (though the numerical values change), so standard sparse
factorization software can be used to obtain the steps. For particular sparsity structures (for
example, bandedness in the matrices A and G), efficient customized solvers for the linear
system arising at each interior-point iteration can be devised.
For very large problems, interior-point methods are often more efficient. However,
when an estimate of the solution is available (a “warm start”), the active-set approach may
converge rapidly in just a few iterations, particularly if the initial value of x is feasible.
Interior-point methods are less able to exploit a warm start, though research efforts to
improve their performance in this regard are ongoing.
16.7
THE GRADIENT PROJECTION METHOD
In the active-set method described in Section 16.5, the active set and working set change
slowly, usually by a single index at each iteration. This method may thus require many
iterations to converge on large-scale problems. For instance, if the starting point x 0 has no
active constraints, while 200 constraints are active at the (nondegenerate) solution, then at
least 200 iterations of the active-set method will be required to reach the solution.
The gradient projection method allows the active set to change rapidly from iteration
to iteration. It is most efficient when the constraints are simple in form—in particular,
when there are only bounds on the variables. Accordingly, we restrict our attention to the
following bound-constrained problem:
min
x
subject to
q(x) 12 x T Gx + x T c
(16.68a)
l ≤ x ≤ u,
(16.68b)
where G is symmetric and l and u are vectors of lower and upper bounds on the components
of x. We do not make any positive definiteness assumptions on G in this section, because the
gradient projection approach can be applied to both convex and nonconvex problems. The
485
486
CHAPTER 16.
QUADRATIC PROGRAMMING
feasible region defined by (16.68b) is sometimes called a “box” because of its rectangular
shape. Some components of x may lack an upper or a lower bound; we handle these cases
formally by setting the appropriate components of l and u to −∞ and +∞, respectively.
Each iteration of the gradient projection algorithm consists of two stages. In the first
stage, we search along the steepest descent direction from the current point x, that is, the
direction −g, where g Gx + c; see (16.6). Whenever a bound is encountered, the search
direction is “bent” so that it stays feasible. We search along the resulting piecewise-linear
path and locate the first local minimizer of q, which we denote by x c and refer to as the
Cauchy point, by analogy with our terminology of Chapter 4. The working set is now defined
to be the set of bound constraints that are active at the Cauchy point, denoted by A(x c ). In
the second stage of each gradient projection iteration, we explore the face of the feasible box
on which the Cauchy point lies by solving a subproblem in which the active components xi
for i ∈ A(x c ) are fixed at the values xic .
We describe the gradient projection method in detail in the rest of this section. Our
convention in this section is to denote the iteration number by a superscript (that is, x k )
and use subscripts to denote the elements of a vector.
CAUCHY POINT COMPUTATION
We now derive an explicit expression for the piecewise-linear path obtained by projecting the steepest descent direction onto the feasible box, and outline the search procedure
for identifying the first local minimum of q along this path.
The projection of an arbitrary point x onto the feasible region (16.68b) is defined as
follows. The ith component is given by
⎧
⎪
l
⎪
⎨ i
xi
P(x, l, u)i ⎪
⎪
⎩ ui
if
x i < li ,
if
xi ∈ [li , u i ],
if
xi > u i .
(16.69)
(We assume, without loss of generality, that li < u i for all i.) The piecewise-linear path x(t)
starting at the reference point x and obtained by projecting the steepest descent direction at
x onto the feasible region (16.68b) is thus given by
x(t) P(x − tg, l, u),
(16.70)
where g Gx + c; see Figure 16.4.
The Cauchy point x c , is defined as the first local minimizer of the univariate, piecewisequadratic function q(x(t)), for t ≥ 0. This minimizer is obtained by examining each of the
line segments that make up x(t). To perform this search, we need to determine the values of
t at which the kinks in x(t), or breakpoints, occur. We first identify the values of t for which
each component reaches its bound along the chosen direction −g. These values t¯i are given
16.7.
THE GRADIENT PROJECTION METHOD
x−tg
x(t3 )
x(t2 )
x(t1 )
x
Figure 16.4
The piecewise-linear path x(t), for an example in IR3 .
by the following explicit formulae:
⎧
⎪
⎨ (xi − u i )/gi
t¯i (xi − li )/gi
⎪
⎩
∞
if gi < 0 and u i < +∞,
if gi > 0 and li > −∞,
(16.71)
otherwise.
The components of x(t) for any t are therefore
xi (t) xi − tgi
if t ≤ t¯i ,
xi − t¯i gi
otherwise.
To search for the first local minimizer along P(x −tg, l, u), we eliminate the duplicate
values and zero values of t¯i from the set {t¯1 , t¯2 , . . . , t¯n }, to obtain a sorted, reduced set
of breakpoints {t1 , t2 , . . . , tl } with 0 < t1 < t2 < · · ·. We now examine the intervals
[0, t1 ], [t1 , t2 ], [t2 , t3 ], . . . in turn. Suppose we have examined up to t j−1 and have not yet
found a local minimizer. For the interval [t j−1 , t j ], we have that
x(t) x(t j−1 ) + (t) p j−1 ,
where
t t − t j−1 ∈ [0, t j − t j−1 ],
487
488
CHAPTER 16.
QUADRATIC PROGRAMMING
and
j−1
pi
−gi
if t j−1 < t¯i ,
0
otherwise.
(16.72)
We can then write the quadratic (16.68a) on the line segment [x(t j−1 ), x(t j )] as
follows:
q(x(t)) c T (x(t j−1 ) + (t) p j−1 ) + 12 (x(t j−1 ) + (t) p j−1 )T G(x(t j−1 ) + (t) p j−1 ).
Expanding and grouping the coefficients of 1, t, and (t)2 , we find that
q(x(t)) f j−1 + f j−1
t +
1
2
f j−1
(t)2 ,
t ∈ [0, t j − t j−1 ],
(16.73)
where the coefficients f j−1 , f j−1
, and f j−1
are defined by
def
f j−1 c T x(t j−1 ) + 12 x(t j−1 )T Gx(t j−1 ),
f j−1
c T p j−1 + x(t j−1 )T Gp j−1 ,
def
f j−1
( p j−1 )T Gp j−1 .
def
Differentiating (16.73) with respect to t and equating to zero, we obtain t ∗ / f j−1
. The following cases can occur. (i) If f j−1
> 0 there is a local minimizer
− f j−1
∗
of q(x(t)) at t t j−1 ; else (ii) t ∈ [0, t j − t j−1 ) there is a minimizer at t t j−1 + t ∗ ;
(iii) in all other cases we move on to the next interval [t j , t j+1 ] and continue the
search.
For the next search interval, we need to calculate the new direction p j from (16.72),
and we use this new value to calculate f j , f j , and f j . Since p j differs from p j−1 typically
in just one component, computational savings can be made by updating these coefficients
rather than computing them from scratch.
SUBSPACE MINIMIZATION
After the Cauchy point x c has been computed, the components of x c that are at their
lower or upper bounds define the active set
A(x c ) {i | xic li or xic u i }.
In the second stage of the gradient projection iteration, we approximately solve the QP
obtained by fixing the components xi for i ∈ A(x c ) at the values xic . The remaining
16.7.
THE GRADIENT PROJECTION METHOD
components are determined from the subproblem
min q(x) 12 x T Gx + x T c
(16.74a)
x
xi xic , i ∈ A(x c ),
(16.74b)
/ A(x ).
li ≤ x i ≤ u i , i ∈
(16.74c)
subject to
c
It is not necessary to solve this problem exactly. Nor is it desirable in the large-dimensional
case, because the subproblem may be almost as difficult as the original problem (16.68).
In fact, to obtain global convergence of the gradient projection procedure, we require only
that the approximate solution x + of (16.74) is feasible with respect to (16.68b) and has an
objective function value no worse than that of x c , that is, q(x + ) ≤ q(x c ). A strategy that is
intermediate between choosing x + x c as the approximate solution (on the one hand) and
solving (16.74) exactly (on the other hand) is to compute an approximate solution of (16.74)
by using the conjugate gradient iteration described in Algorithm 16.1 or Algorithm 16.2.
Note that for the equality constraints (16.74b), the Jacobian A and the null-space basis
matrix Z have particularly simple forms. We could therefore apply conjugate gradient to
the problem (16.74a), (16.74b) and terminate as soon as a bound l ≤ x ≤ u is encountered.
Alternatively, we could continue to iterate, temporarily ignoring the bounds and projecting
the solution back onto the box constraints. The negative-curvature case can be handled
as in Algorithm 7.2, the method for approximately solving possibly indefinite trust-region
subproblems in unconstrained optimization.
We summarize the gradient projection algorithm for quadratic programming as
follows.
Algorithm 16.5 (Gradient Projection Method for QP).
Compute a feasible starting point x 0 ;
for k 0, 1, 2, . . .
if x k satisfies the KKT conditions for (16.68)
stop with solution x ∗ x k ;
Set x x k and find the Cauchy point x c ;
Find an approximate solution x + of (16.74) such that q(x + ) ≤ q(x c )
and x + is feasible;
k+1
← x +;
x
end (for)
If the algorithm approaches a solution x ∗ at which the Lagrange multipliers associated
with all the active bounds are nonzero (that is, strict complementarity holds), the active sets
A(x c ) generated by the gradient projection algorithm are equal to the optimal active set for
all k sufficiently large. That is, constraint indices do not repeatedly enter and leave the active
set on successive iterations. When the problem is degenerate, the active set may not settle
down at its optimal value. Various devices have been proposed to prevent this undesirable
behavior from taking place.
489
490
CHAPTER 16.
QUADRATIC PROGRAMMING
While gradient projection methods can be applied in principle to problems with general linear constraints, significant computation may be required to perform the projection
onto the feasible set in such cases. For example, if the constraint set is defined as aiT x ≥ bi ,
i ∈ I, we must solve the following convex quadratic program to compute the projection of
a given point x̄ onto this set:
max x − x̄2
x
subject to aiT x ≥ bi for all i ∈ I.
The expense of solving this “projection subproblem” may approach the cost of solving the
original quadratic program, so it is usually not economical to apply gradient projection to
this case.
When we use duality to replace a strictly convex quadratic program with its dual
(see Example 12.12), the gradient projection method may be useful in solving the boundconstrained dual problem, which is formulated in terms of the Lagrange multipliers λ as
follows:
1
max q̃(λ) − (A T λ − c)T G −1 (A T λ − c)T + b T λ, subject to λ ≥ 0.
λ
2
(Note that the dual is conventionally written as a maximization problem; we can equivalently
minimize −q̃(λ) and note that this transformed problem is convex.) This approach is most
useful when G has a simple form, for example, a diagonal or block-diagonal matrix.
16.8
PERSPECTIVES AND SOFTWARE
Active-set methods for convex quadratic programming are implemented in QPOPT [126],
VE09 [142], BQPD [103], and QPA [148]. Several commercial interior-point solvers for QP
are available, including CPLEX [172], XPRESS-MP [159] and MOSEK [5]. The code QPB [146]
uses a two-phase interior-point method that can handle convex and nonconvex problems.
OOPS [139] and OOQP [121] are object-oriented interior-point codes that allow the user
to customize the linear algebra techniques to the particular structure of the data for an
application. Some nonlinear programming interior-point packages, such as LOQO [294] and
KNITRO [46], are also effective for convex and nonconvex quadratic programming.
The numerical comparison of active-set and interior-point methods for convex
quadratic programming reported in [149] indicates that interior-point methods are generally much faster on large problems. If a warm start is required, however, active-set methods
may be generally preferable. Although considerable research has been focused on improving
the warm-start capabilities of interior-point methods, the full potential of such techniques
is now yet known.
We have assumed in this chapter that all equality-constrained quadratic programs have
linearly independent constraints, that is, the m × n constraint Jacobian matrix A has rank m.
16.8.
PERSPECTIVES
AND
SOFTWARE
If redundant constraints are present, they can be detected by forming a SVD or rank-revealing
QR factorization of A T , and then removed from the formulation. When A is larger, sparse
Gaussian elimination techniques can be applied to A T instead, but they are less reliable.
The KNITRO and OOPS software packages provide the option of solving the primal-dual
equations (16.63) by means of the projected CG iteration of Algorithm 16.2.
We have not considered active-set methods for the case in which the Hessian matrix
G is indefinite because these methods can be quite complicated to describe and it is not well
understood how to adapt them to the large dimensional case. We make some comments
here on the principal techniques.
Algorithm 16.3, the active-set method for convex QP, can be adapted to this indefinite
case by modifying the computation of the search direction and step length in certain
situations. To explain the need for the modification, we consider the computation of a
step by a null-space method, that is, p Z pZ , where pZ is given by (16.53). If the reduced
Hessian Z T G Z is positive definite, then this step p points to the minimizer of the subproblem
(16.39), and the logic of the iteration need not be changed. If Z T G Z has negative eigenvalues,
however, p points only to a saddle point of (16.39) and is therefore not always a suitable
step. Instead, we seek an alternative direction sZ that is a direction of negative curvature for
Z T G Z . We then have that
q(x + α Z sZ ) → −∞
as α → ∞.
(16.75)
Additionally, we change the sign of sZ if necessary to ensure that Z sZ is a non-ascent direction
for q at the current point x, that is, ∇q(x)T Z sZ ≤ 0. By moving along the direction Z sZ , we
will encounter a constraint that can be added to the working set for the next iteration. (If we
don’t find such a constraint, the problem is unbounded.) If the reduced Hessian for the new
working set is not positive definite, we repeat this process until enough constraints have been
added to make the reduced Hessian positive definite. A difficulty with this general approach,
however, is that if we allow the reduced Hessian to have several negative eigenvalues, it
is difficult to make these methods efficient when the reduced Hessian changes from one
working set to the next.
Inertia controlling methods are a practical class of algorithms for indefinite QP that
never allow the reduced Hessian to have more than one negative eigenvalue. As in the convex
case, there is a preliminary phase in which a feasible starting point x0 is found. We place
the additional demand on x0 that it be either a vertex (in which case the reduced Hessian is
the null matrix) or a constrained stationary point at which the reduced Hessian is positive
definite. At each iteration, the algorithm will either add or remove a constraint from the
working set. If a constraint is added, the reduced Hessian is of smaller dimension and must
remain positive definite or be the null matrix. Therefore, an indefinite reduced Hessian can
arise only when one of the constraints is removed from the working set, which happens
only when the current point is a minimizer with respect to the current working set. In this
case, we will choose the new search direction to be a direction of negative curvature for the
reduced Hessian.
491
492
CHAPTER 16.
QUADRATIC PROGRAMMING
Various algorithms for indefinite QP differ in the way that indefiniteness is detected,
in the computation of the negative curvature direction, and in the handling of the working
set; see Fletcher [99] and Gill and Murray [126].
NOTES AND REFERENCES
The problem of determining whether a feasible point for a nonconvex QP (16.1) is a
global minimizer is NP-hard (Murty and Kabadi [219]); so is the problem of determining
whether a given point is a local minimizer (Vavasis [296, Theorem 5.1]). Various algorithms
for convex QP with polynomial convexity are discussed in Nesterov and Nemirovskii [226].
The portfolio optimization problem was formulated by Markowitz [201].
For a discussion on the QMR, LSQR, and GMRES methods see, for example, [136,
272, 290]. The idea of using the projection (16.30) in the CG method dates back to at
least Polyak [238]. The alternative (16.34), and its special case (16.32), are proposed in
Coleman [64]. Although it can give rise to substantial rounding errors, they can be corrected
by iterative refinement; see Gould et al. [143]. More recent studies on preconditioning of
the projected CG method include Keller et al. [176] and Lukšan and Vlček [196].
For further discussion on the gradient projection method see, for example, Conn,
Gould, and Toint [70] and Burke and Moré [44].
In some areas of application, the KKT matrix (16.7) not only is sparse but also contains
special structure. For instance, the quadratic programs that arise in many control problems
have banded matrices G and A (see Wright [315]), which can be exploited by interior-point
methods via a suitable symmetric reordering of K . When active-set methods are applied to
this problem, however, the advantages of bandedness and sparsity are lost after just a few
updates of the factorization.
Further details of interior-point methods for convex quadratic programming can
be found in Wright [316] and Vanderbei [293]. The first inertia-controlling method for
indefinite quadratic programming was proposed by Fletcher [99]. See also Gill et al. [129]
and Gould [142] for a discussion of methods for general quadratic programming.
✐
EXERCISES
✐
16.1
(a) Solve the following quadratic program and illustrate it geometrically.
min f (x) 2x1 + 3x2 + 4x12 + 2x1 x2 + x22 ,
subject to x1 − x2 ≥ 0,
x1 + x2 ≤ 4,
x1 ≤ 3.
(b) If the objective function is redefined as q(x) − f (x), does the problem have a finite
minimum? Are there local minimizers?
16.8.
PERSPECTIVES
AND
SOFTWARE
✐
16.2 The problem of finding the shortest distance from a point x0 to the hyperplane
{x | Ax b}, where A has full row rank, can be formulated as the quadratic program
min 12 (x − x0 )T (x − x0 ) subject to Ax b.
Show that the optimal multiplier is
λ∗ (A A T )−1 (b − Ax0 )
and that the solution is
x ∗ x0 + A T (A A T )−1 (b − Ax0 ).
Show that in the special case in which A is a row vector, the shortest distance from x0 to the
solution set of Ax b is |b − Ax0 |/A2 .
✐
16.3 Use Theorem 12.1 to verify that the first-order necessary conditions for (16.3)
are given by (16.4).
✐
16.4 Suppose that G is positive semidefinite in (16.1) and that x ∗ satisfies the KKT
conditions (16.37) for some λi∗ , i ∈ A(x ∗ ). Suppose in addition that second-order sufficient
conditions are satisfied, that is, Z T G Z is positive definite where the columns of Z span the
null space of the active constraint Jacobian matrix. Show that x ∗ is in fact the unique global
solution for (16.1), that is, q(x) > q(x ∗ ) for all feasible x with x x ∗ .
✐
16.5 Verify that the inverse of the KKT matrix is given by (16.16).
✐
16.6 Use Theorem 12.6 to show that if the conditions of Lemma 16.1 hold, then the
second-order sufficient conditions for (16.3) are satisfied by the vector pair (x ∗ , λ∗ ) that
satisfies (16.4).
✐
16.7 Consider (16.3) and suppose that the projected Hessian matrix Z T G Z has a
negative eigenvalue; that is, u T Z T G Z u < 0 for some vector u. Show that if there exists any
vector pair (x ∗ , λ∗ ) that satisfies (16.4), then the point x ∗ is only a stationary point of (16.3)
and not a local minimizer. (Hint: Consider the function q(x ∗ + α Z u) for α 0, and use
an expansion like that in the proof of Theorem 16.2.)
✐
16.8 By using the QR factorization and a permutation matrix, show that for a fullrank m ×n matrix A (with m < n) one)can find an
* orthogonal matrix Q and an m ×m upper
triangular matrix Û such that AQ 0 Û . (Hint: Start by applying the standard QR
factorization to A T .)
✐
16.9 Verify that the first-order conditions for optimality of (16.1) are equivalent to
(16.37) when we make use of the active-set definition (16.36).
493
494
CHAPTER 16.
QUADRATIC PROGRAMMING
✐
16.10 For each of the alternative choices of initial working set W0 in the example
(16.49) (that is, W0 {3}, W0 {5}, and W0 ∅) work through the first two iterations of
Algorithm 16.3.
✐
16.11 Program Algorithm 16.3, and use it to solve the problem
min
subject to
x12 + 2x22 − 2x1 − 6x2 − 2x1 x2
1
x
2 1
+ 12 x2 ≤ 1,
−x1 + 2x2 ≤ 2,
x1 , x2 ≥ 0.
Choose three initial starting points: one in the interior of the feasible region, one at a vertex,
and one at a non-vertex point on the boundary of the feasible region.
✐ 16.12 Show that the operator P defined by (16.27) is independent of the choice of
null-space basis Z . (Hint: First show that any null-space basis Z can be written as Z Q B
where Q is an orthogonal basis and B is a nonsingular matrix.)
✐
16.13
(a) Show that the the computation of the preconditioned residual g + in (16.28d) can be
performed with (16.29) or (16.30).
(b) Show that we can also perform this computation by solving the system (16.32).
(c) Verify (16.33).
✐
16.14
(a) Show that if Z T G Z is positive definite, then the denominator in (16.28a) is nonzero.
(b) Show that if Z T r rZ 0 and Z T H Z is positive definite, then the denominator in
(16.28e) is nonzero.
✐
16.15 Consider problem (16.3), and assume that A has full row rank and that Z is a
basis for the null space of A. Prove that there are no finite solutions if Z T G Z has negative
eigenvalues.
✐
16.16
(a) Assume that A 0. Show that the KKT matrix (16.7) is indefinite.
(b) Prove that if the KKT matrix (16.7) is nonsingular, then A must have full rank.
✐
16.17 Consider the quadratic program
max
subject to
6x1 + 4x2 − 13 − x12 − x22 ,
x1 + x2 ≤ 3,
x1 ≥ 0,
x2 ≥ 0.
(16.76)
16.8.
PERSPECTIVES
AND
SOFTWARE
First solve it graphically, and then use your program implementing the active-set method
given in Algorithm 16.3.
✐
16.18 Using (16.39) and (16.41), explain briefly why the gradient of each blocking
constraint cannot be a linear combination of the constraint gradients in the current working
set Wk .
✐
16.19 Let W be an n × n symmetric matrix, and suppose that Z is of dimension n × t.
Suppose that Z T W Z is positive definite and that Z̄ is obtained by removing a column from
Z . Show that Z̄ T W Z̄ is positive definite.
✐ 16.20 Find a null-space basis matrix Z for the equality-constrained problem defined
by (16.74a), (16.74b).
✐
16.21 Write down KKT conditions for the following convex quadratic program with
mixed equality and inequality constraints:
min q(x) 12 x T Gx + x T c
subject to
Ax ≥ b, Āx b̄,
where G is symmetric and positive semidefinite. Use these conditions to derive an analogue
of the generic primal-dual step (16.58) for this problem.
✐
16.22 Explain why for a bound-constrained problems the number of possible active
sets is at most 3n .
✐
16.23
(a) Show that the primal-dual system (16.58) can be solved using the augmented system
(16.61) or the normal equations (16.62). Describe in detail how all the components
(x, y, λ) are computed.
(b) Verify (16.65).
✐
16.24 Program Algorithm 16.4 and use it to solve problem (16.76). Set all initial
variables to be the vector e (1, 1, . . . , 1)T .
✐
16.25 Let x̄ ∈ R n be given, and let x ∗ be the solution of the projection problem
min x − x̄2
subject to
l ≤ x ≤ u.
(16.77)
For simplicity, assume that −∞ < li < u i < ∞ for all i 1, 2, . . . , n. Show that the
solution of this problem coincides with the projection formula given by (16.69) that is, show
that x ∗ P(x̄, l, u). (Hint: Note that the problem is separable.)
495
496
CHAPTER 16.
✐
QUADRATIC PROGRAMMING
16.26 Consider the bound-constrained quadratic problem (16.68) with
G
4
1
1
2
, c
−1
1
, l
0
0
, and u 5
3
.
(16.78)
Suppose x 0 (0, 2)T . Find t¯1 , t¯2 , t1 , t2 , p 1 , p 2 and x(t1 ), x(t2 ). Find the minimizer of
q(x(t)).
✐
16.27 Consider the search for the one dimensional minimizer of the function q(x(t))
defined by (16.73). There are 9 possible cases since f, f , f can each be positive, negative, or
zero. For each case, determine the location of the minimizer. Verify that the rules described
in Section 16.7 hold.
This is page 497
Printer: Opaque this
17
CHAPTER
Penalty and
Augmented
Lagrangian
Methods
Some important methods for constrained optimization replace the original problem by a
sequence of subproblems in which the constraints are represented by terms added to the
objective. In this chapter we describe three approaches of this type. The quadratic penalty
method adds a multiple of the square of the violation of each constraint to the objective.
Because of its simplicity and intuitive appeal, this approach is used often in practice, although it has some important disadvantages. In nonsmooth exact penalty methods, a single
unconstrained problem (rather than a sequence) takes the place of the original constrained
problem. Using these penalty functions, we can often find a solution by performing a single
498
CHAPTER 17.
PENALTY
AND
AUGMENTED LAGRANGIAN METHODS
unconstrained minimization, but the nonsmoothness may create complications. A popular
function of this type is the 1 penalty function. A different kind of exact penalty approach
is the method of multipliers or augmented Lagrangian method, in which explicit Lagrange
multiplier estimates are used to avoid the ill-conditioning that is inherent in the quadratic
penalty function.
A somewhat related approach is used in the log-barrier method, in which logarithmic
terms prevent feasible iterates from moving too close to the boundary of the feasible region. This approach forms part of the foundation for interior-point methods for nonlinear
programming and we discuss it further in Chapter 19.
17.1
THE QUADRATIC PENALTY METHOD
MOTIVATION
Let us consider replacing a constrained optimization problem by a single function
consisting of
- the original objective of the constrained optimization problem, plus
- one additional term for each constraint, which is positive when the current point x
violates that constraint and zero otherwise.
Most approaches define a sequence of such penalty functions, in which the penalty terms for
the constraint violations are multiplied by a positive coefficient. By making this coefficient
larger, we penalize constraint violations more severely, thereby forcing the minimizer of the
penalty function closer to the feasible region for the constrained problem.
The simplest penalty function of this type is the quadratic penalty function, in which
the penalty terms are the squares of the constraint violations. We describe this approach
first in the context of the equality-constrained problem
min f (x)
x
subject to ci (x) 0, i ∈ E,
(17.1)
which is a special case of (12.1). The quadratic penalty function Q(x; µ) for this formulation
is
def
Q(x; µ) f (x) +
µ 2
c (x),
2 i∈E i
(17.2)
where µ > 0 is the penalty parameter. By driving µ to ∞, we penalize the constraint
violations with increasing severity. It makes good intuitive sense to consider a sequence of
values {µk } with µk ↑ ∞ as k → ∞, and to seek the approximate minimizer xk of Q(x; µk )
for each k. Because the penalty terms in (17.2) are smooth, we can use techniques from
17.1.
THE QUADRATIC PENALTY METHOD
unconstrained optimization to search for xk . In searching for xk , we can use the minimizers
xk−1 , xk−2 , etc., of Q(·; µ) for smaller values of µ to construct an initial guess. For suitable
choices of the sequence {µk } and the initial guesses, just a few steps of unconstrained
minimization may be needed for each µk .
❏ EXAMPLE 17.1
Consider the problem (12.9) from Chapter 12, that is,
min x1 + x2
subject to x12 + x22 − 2 0,
(17.3)
for which the solution is (−1, −1)T and the quadratic penalty function is
Q(x; µ) x1 + x2 +
2
µ 2
x1 + x22 − 2 .
2
(17.4)
We plot the contours of this function in Figures 17.1 and 17.2. In Figure 17.1 we have
µ 1, and we observe a minimizer of Q near the point (−1.1, −1.1)T . (There is also a
local maximizer near x (0.3, 0.3)T .) In Figure 17.2 we have µ 10, so points that do not
lie on the feasible circle defined by x12 + x22 2 suffer a much greater penalty than in the
first figure—the “trough” of low values of Q is clearly evident. The minimizer in this figure
is much closer to the solution (−1, −1)T of the problem (17.3). A local maximum lies near
(0, 0)T , and Q goes rapidly to ∞ outside the circle x12 + x22 2.
❐
1.5
1
0.5
0
−0.5
−1
−1.5
−1.5
−1
−0.5
0
0.5
1
1.5
Figure 17.1 Contours of Q(x; µ) from (17.4) for µ 1, contour spacing 0.5.
499
500
CHAPTER 17.
PENALTY
AND
AUGMENTED LAGRANGIAN METHODS
1.5
1
0.5
0
−0.5
−1
−1.5
−1.5
−1
Figure 17.2
−0.5
0
0.5
1
1.5
Contours of Q(x; µ) from (17.4) for µ 10, contour spacing 2.
The situation is not always so benign as in Example 17.1. For a given value of the
penalty parameter µ, the penalty function may be unbounded below even if the original
constrained problem has a unique solution. Consider for example
min −5x12 + x22
subject to x1 1,
(17.5)
whose solution is (1, 0)T . The penalty function is unbounded for any µ < 10. For such
values of µ, the iterates generated by an unconstrained minimization method would usually
diverge. This deficiency is, unfortunately, common to all the penalty functions discussed in
this chapter.
For the general constrained optimization problem
min f (x)
x
subject to ci (x) 0, i ∈ E,
ci (x) ≥ 0,
i ∈ I,
(17.6)
which contains inequality constraints as well as equality constraints, we can define the
quadratic penalty function as
def
Q(x; µ) f (x) +
2
µ 2
µ [ci (x)]− ,
ci (x) +
2 i∈E
2 i∈I
(17.7)
where [y]− denotes max(−y, 0). In this case, Q may be less smooth than the objective and
constraint functions. For instance, if one of the inequality constraints is x1 ≥ 0, then the
function min(0, x1 )2 has a discontinuous second derivative, so that Q is no longer twice
continuously differentiable.
17.1.
THE QUADRATIC PENALTY METHOD
ALGORITHMIC FRAMEWORK
A general framework for algorithms based on the quadratic penalty function (17.2)
can be specified as follows.
Framework 17.1 (Quadratic Penalty Method).
Given µ0 > 0, a nonnegative sequence {τk } with τk → 0, and a starting point x0s ;
for k 0, 1, 2, . . .
Find an approximate minimizer xk of Q(·; µk ), starting at xks ,
and terminating when ∇x Q(x; µk ) ≤ τk ;
if final convergence test satisfied
stop with approximate solution xk ;
end (if)
Choose new penalty parameter µk+1 > µk ;
s
;
Choose new starting point xk+1
end (for)
The parameter sequence {µk } can be chosen adaptively, based on the difficulty of
minimizing the penalty function at each iteration. When minimization of Q(x; µk ) proves
to be expensive for some k, we choose µk+1 to be only modestly larger than µk ; for instance
µk+1 1.5µk . If we find the approximate minimizer of Q(x; µk ) cheaply, we could try a
more ambitious increase, for instance µk+1 10µk . The convergence theory for Framework 17.1 allows wide latitude in the choice of nonnegative tolerances τk ; it requires only
that τk → 0, to ensure that the minimization is carried out more accurately as the iterations
progress.
There is no guarantee that the stop test ∇x Q(x; µk ) ≤ τk will be satisfied because, as discussed above, the iterates may move away from the feasible region when the
penalty parameter is not large enough. A practical implementation must include safeguards that increase the penalty parameter (and possibly restore the initial point) when
the constraint violation is not decreasing rapidly enough, or when the iterates appear to be
diverging.
When only equality constraints are present, Q(x; µk ) is smooth, so the algorithms for
unconstrained minimization described in the first chapters of the book can be used to identify
the approximate solution xk . However, the minimization of Q(x; µk ) becomes more difficult
to perform as µk becomes large, unless we use special techniques to calculate the search
directions. For one thing, the Hessian ∇x2x Q(x; µk ) becomes arbitrarily ill conditioned near
the minimizer. This property alone is enough to make many unconstrained minimization
algorithms such as quasi-Newton and conjugate gradient perform poorly. Newton’s method,
on the other hand, is not sensitive to ill conditioning of the Hessian, but it, too, may encounter
difficulties for large µk for two other reasons. First, ill conditioning of ∇x2x Q(x; µk ) might
be expected to cause numerical problems when we solve the linear equations to calculate
the Newton step. We discuss this issue below, and show that these effects are not severe and
501
502
CHAPTER 17.
PENALTY
AND
AUGMENTED LAGRANGIAN METHODS
that a reformulation of the Newton equations is possible. Second, even when x is close to
the minimizer of Q(·; µk ), the quadratic Taylor series approximation to Q(x; µk ) about
x is a reasonable approximation of the true function only in a small neighborhood of x.
This property can be seen in Figure 17.2, where the contours of Q near the minimizer have
a “banana” shape, rather than the elliptical shape that characterizes quadratic functions.
Since Newton’s method is based on the quadratic model, the steps that it generates may not
make rapid progress toward the minimizer of Q(x; µk ). This difficulty can be lessened by a
s
s
, or by setting xk+1
xk and choosing µk+1 to be
judicious choice of the starting point xk+1
only modestly larger than µk .
CONVERGENCE OF THE QUADRATIC PENALTY METHOD
We describe some convergence properties of the quadratic penalty method in the
following two theorems. We restrict our attention to the equality-constrained problem
(17.1), for which the quadratic penalty function is defined by (17.2).
For the first result we assume that the penalty function Q(x; µk ) has a (finite)
minimizer for each value of µk .
Theorem 17.1.
Suppose that each xk is the exact global minimizer of Q(x; µk ) defined by (17.2) in
Framework 17.1 above, and that µk ↑ ∞. Then every limit point x ∗ of the sequence {xk } is a
global solution of the problem (17.1).
PROOF. Let x̄ be a global solution of (17.1), that is,
f (x̄) ≤ f (x)
for all x with ci (x) 0,
i ∈ E.
Since xk minimizes Q(·; µk ) for each k, we have that Q(xk ; µk ) ≤ Q(x̄; µk ), which leads
to the inequality
f (xk ) +
µk 2
µk 2
ci (xk ) ≤ f (x̄) +
c (x̄) f (x̄).
2 i∈E
2 i∈E i
(17.8)
By rearranging this expression, we obtain
ci2 (xk ) ≤
i∈E
2
[ f (x̄) − f (xk )].
µk
(17.9)
Suppose that x ∗ is a limit point of {xk }, so that there is an infinite subsequence K such that
lim xk x ∗ .
k∈K
17.1.
THE QUADRATIC PENALTY METHOD
By taking the limit as k → ∞, k ∈ K, on both sides of (17.9), we obtain
i∈E
ci2 (x ∗ ) lim
k∈K
2
[ f (x̄) − f (xk )] 0,
k∈K µk
ci2 (xk ) ≤ lim
i∈E
where the last equality follows from µk ↑ ∞. Therefore, we have that ci (x ∗ ) 0 for all
i ∈ E, so that x ∗ is feasible. Moreover, by taking the limit as k → ∞ for k ∈ K in (17.8), we
have by nonnegativity of µk and of each ci (xk )2 that
µk 2
ci (xk ) ≤ f (x̄).
k∈K 2
i∈E
f (x ∗ ) ≤ f (x ∗ ) + lim
Since x ∗ is a feasible point whose objective value is no larger than that of the global solution
x̄, we conclude that x ∗ , too, is a global solution, as claimed.
Since this result requires us to find the global minimizer for each subproblem, this
desirable property of convergence to the global solution of (17.1) cannot be attained in
general. The next result concerns convergence properties of the sequence {xk } when we allow
inexact (but increasingly accurate) minimizations of Q(·; µk ). In contrast to Theorem 17.1,
it shows that the sequence may be attracted to infeasible points, or to any KKT point (that is,
a point satisfying first-order necessary conditions; see (12.34)), rather than to a minimizer. It
also shows that the quantities µk ci (xk ) may be used as estimates of the Lagrange multipliers
λi∗ in certain circumstances. This observation is important for the analysis of augmented
Lagrangian methods in Section 17.3.
To establish the result we will make the (optimistic) assumption that the stop test
∇x Q(x; µk ) ≤ τk is satisfied for all k.
Theorem 17.2.
Suppose that the tolerances and penalty parameters in Framework 17.1 satisfy τk → 0
and µk ↑ ∞. Then if a limit point x ∗ of the sequence {xk } is infeasible, it is a stationary point
of the function c(x)2 . On the other hand, if a limit point x ∗ is feasible and the constraint
gradients ∇ci (x ∗ ) are linearly independent, then x ∗ is a KKT point for the problem (17.1). For
such points, we have for any infinite subsequence K such that limk∈K xk x ∗ that
lim −µk ci (xk ) λi∗ ,
for all i ∈ E,
k∈K
(17.10)
where λ∗ is the multiplier vector that satisfies the KKT conditions (12.34) for the equalityconstrained problem (17.1).
PROOF. By differentiating Q(x; µk ) in (17.2), we obtain
∇x Q(xk ; µk ) ∇ f (xk ) +
i∈E
µk ci (xk )∇ci (xk ),
(17.11)
503
504
CHAPTER 17.
PENALTY
AND
AUGMENTED LAGRANGIAN METHODS
so from the termination criterion for Framework 17.1, we have that
µk ci (xk )∇ci (xk ) ≤ τk .
∇ f (xk ) +
(17.12)
i∈E
By rearranging this expression (and in particular using the inequality a−b ≤ a +b),
we obtain
1
ci (xk )∇ci (xk ) ≤
(17.13)
[τk + ∇ f (xk )] .
µk
i∈E
Let x ∗ be a limit point of the sequence of iterates. Then there is a subsequence K such
that limk∈K xk x ∗ . When we take limits as k → ∞ for k ∈ K, the bracketed term on the
right-hand-side approaches ∇ f (x ∗ ), so because µk ↑ ∞, the right-hand-side approaches
zero. From the corresponding limit on the left-hand-side, we obtain
ci (x ∗ )∇ci (x ∗ ) 0.
(17.14)
i∈E
We can have ci (x ∗ ) 0 (if the constraint gradients ∇ci (x ∗ ) are dependent), but in this case
(17.14) implies that x ∗ is a stationary point of the function c(x)2 .
If, on the other hand, the constraint gradients ∇ci (x ∗ ) are linearly independent at a
limit point x ∗ , we have from (17.14) that ci (x ∗ ) 0 for all i ∈ E, so x ∗ is feasible. Hence,
the second KKT condition (12.34b) is satisfied. We need to check the first KKT condition
(12.34a) as well, and to show that the limit (17.10) holds.
By using A(x) to denote the matrix of constraint gradients (also known as the
Jacobian), that is,
A(x)T [∇ci (x)]i∈E ,
(17.15)
and λk to denote the vector −µk c(xk ), we have as in (17.12) that
A(xk )T λk ∇ f (xk ) − ∇x Q(xk ; µk ),
∇x Q(xk ; µk ) ≤ τk .
(17.16)
For all k ∈ K sufficiently large, the matrix A(xk ) has full row rank, so that A(xk )A(xk )T is
nonsingular. By multiplying (17.16) by A(xk ) and rearranging, we have that
λk A(xk )A(xk )T
−1
A(xk ) [∇ f (xk ) − ∇x Q(xk ; µk )] .
Hence by taking the limit as k ∈ K goes to ∞, we find that
lim λk λ∗ A(x ∗ )A(x ∗ )T
k∈K
−1
A(x ∗ )∇ f (x ∗ ).
17.1.
THE QUADRATIC PENALTY METHOD
By taking limits in (17.12), we conclude that
∇ f (x ∗ ) − A(x ∗ )T λ∗ 0,
(17.17)
so that λ∗ satisfies the first KKT condition (12.34a) for (17.1). Hence, x ∗ is a KKT point for
(17.1), with unique Lagrange multiplier vector λ∗ .
It is reassuring that, if a limit point x ∗ is not feasible, it is at least a stationary point for
the function c(x)2 . Newton-type algorithms can always be attracted to infeasible points of
this type. (We see the same effect in Chapter 11, in our discussion of methods for nonlinear
equations that use the sum-of-squares merit function r (x)2 .) Such methods cannot be
guaranteed to find a root, and can be attracted to a stationary point or minimizer of the
merit function. In the case in which the nonlinear program (17.1) is infeasible, we often
observe convergence of the quadratic-penalty method to stationary points or minimizers of
c(x)2 .
ILL CONDITIONING AND REFORMULATIONS
We now examine the nature of the ill conditioning in the Hessian ∇x2x Q(x; µk ).
An understanding of the properties of this matrix, and the similar Hessians that arise in
other penalty and barrier methods, is essential in choosing effective algorithms for the
minimization problem and for the linear algebra calculations at each iteration.
The Hessian is given by the formula
∇x2x Q(x; µk ) ∇ 2 f (x) +
µk ci (x)∇ 2 ci (x) + µk A(x)T A(x),
(17.18)
i∈E
where we have used the definition (17.15) of A(x). When x is close to the minimizer of
Q(·; µk ) and the conditions of Theorem 17.2 are satisfied, we have from (17.10) that the
sum of the first two terms on the right-hand-side of (17.18) is approximately equal to the
Hessian of the Lagrangian function defined in (12.33). To be specific, we have
∇x2x Q(x; µk ) ≈ ∇x2x L(x, λ∗ ) + µk A(x)T A(x),
(17.19)
when x is close to the minimizer of Q(·; µk ). We see from this expression that ∇x2x Q(x; µk )
is approximately equal to the sum of
- a matrix whose elements are independent of µk (the Lagrangian term), and
- a matrix of rank |E| whose nonzero eigenvalues are of order µk (the second term on
the right-hand side of (17.19)).
The number of constraints |E| is usually smaller than n. In this case, the last term in (17.19) is
singular. The overall matrix has some of its eigenvalues approaching a constant, while others
505
506
CHAPTER 17.
PENALTY
AND
AUGMENTED LAGRANGIAN METHODS
are of order µk . Since µk is approaching ∞, the increasing ill conditioning of ∇x2x Q(x; µk )
is apparent.
One consequence of the ill conditioning is possible inaccuracy in the calculation of
the Newton step for Q(x; µk ), which is obtained by solving the following system:
∇x2x Q(x; µk ) p −∇x Q(x; µk ).
(17.20)
In general, the poor conditioning of this system will lead to significant errors in the computed
value of p, regardless of the computational technique used to solve (17.20). For the same
reason, iterative methods can be expected to perform poorly unless accompanied by a
preconditioning strategy that removes the systematic ill conditioning.
There is an alternative formulation of the equations (17.20) that avoids the ill conditioning due to the final term in (17.18). By introducing a new variable vector ζ defined by
ζ µA(x) p, we see that the vector p that solves (17.20) also satisfies the following system:
⎡
⎣
∇ 2 f (x) +
µk ci (x)∇ 2 ci (x)
A(x)T
i∈E
A(x)
−(1/µk )I
⎤
⎦
p
ζ
−∇x Q(x; µk )
0
.
(17.21)
When x is not too far from the solution x ∗ , the coefficient matrix in this system does not have
large singular values (of order µk ), so the system (17.21) can be viewed as a well conditioned
reformulation of (17.20). We note, however, that neither system may yield a good search
direction p because the coefficients µk ci (x) in the summation term of the upper left block of
(17.21) may be poor approximations to the Lagrange multipliers −λi∗ , even when x is quite
close to the minimizer xk of Q(x; µk ). This fact may cause the quadratic model on which p
is based to be an inadequate model of Q(·; µk ), so the Newton step may be intrinsically an
unsuitable search direction. We discussed possible remedies for this difficulty above, in our
comments following Framework 17.1.
To compute the step via (17.21) involves the solution of a linear system of dimension
n + |E| rather than the system of dimension n given by (17.19). A similar system must
be solved to calculate the sequential quadratic programming (SQP) step (18.6), which is
derived in Chapter 18. In fact, when µk is large, (17.21) can be viewed as a regularization of
the SQP step (18.6